A Generative Adversarial Network-Driven Method for Molten Salt Anti-condensation and Safe Start-up / Shutdown Control
By combining AC-GANs and MPC, differentiated temperature control trajectories for collector tubes in linear Fresnel solar thermal power plants are automatically generated, solving problems such as molten salt freezing, overheating, and thermal stress damage, and improving equipment safety and operational economy.
Patent Information
- Application Number
- CN202611267424.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-20
- Publication Date
- 2026-11-17
AI Technical Summary
In linear Fresnel solar thermal power plants, the molten salt inside the collector tubes faces multiple safety issues during start-up and shutdown, such as freezing, overheating, and thermal stress damage. Traditional control methods struggle to dynamically generate the optimal temperature control trajectory under multiple constraints.
A multi-region condition generation method based on generative adversarial networks (AC-GANs) is adopted, combined with a model predictive controller (MPC), to automatically generate differentiated temperature control trajectories, and to ensure safety through multi-level temperature hard protection logic.
It achieves optimal temperature control under the differentiated thermal characteristics of multiple regions, reduces the risk of freezing and overheating, improves equipment life and operating efficiency, and reduces energy consumption during downtime.
Smart Images

Figure CN123113333A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of solar thermal power generation technology, specifically relating to a method for molten salt anti-condensation and safe start-up / shutdown control based on generative adversarial networks. Background Technology
[0002] Linear Fresnel solar thermal power plants use binary nitrate (60% NaNO3 + 40% KNO3) as the heat transfer and storage medium, which has the advantages of low cost and high heat storage density. Vacuum collector tubes are the core component for converting solar energy into molten salt thermal energy. However, in actual engineering operation, the start-up and shutdown process of the collector tube-molten salt coupling body faces the following technical challenges: 1. Cold Start Thermal Stress Damage. During early morning startup, the low-temperature collector tubes suddenly come into contact with high-temperature molten salt, or the molten salt rapidly heats up from its freezing point (approximately 240°C), generating enormous thermal stress. Fixed temperature rise rate control cannot balance equipment lifespan and startup efficiency.
[0003] 2. Risk of freezing and blockage during shutdown. If the insulation is not properly maintained when the system is shut down at night or on cloudy days, the molten salt inside the heat collection tubes can easily solidify (melting point about 220°C), causing the pipeline to freeze and become paralyzed. Traditional constant temperature heat tracing or constant speed circulation strategies consume a lot of energy.
[0004] 3. Risk of high-temperature decomposition. If not properly controlled during operation, the molten salt outlet temperature can easily approach the upper limit of 550℃, or even trigger multi-level protection at 542℃ / 545℃ / 547℃. Frequent derating or emergency cooling will seriously affect the continuity of power generation and the life of the equipment.
[0005] 4. Significant differences in thermal characteristics across multiple regions. The molten salt pipelines of the linear Fresnel solar thermal power plant exhibit significant differences in thermal inertia, heat dissipation conditions, and safety margins across different regions (main plant, mirror field branch lines, mirror field main line, transport pipeline section between the main plant and mirror field, and valves). Traditional fixed parameter control cannot adaptively adjust these differences, resulting in excessive insulation in some areas leading to energy waste, while insufficient insulation in other areas poses risks of freezing blockage and thermal stress damage.
[0006] Existing PID control or APS systems based on fixed rules struggle to dynamically generate the optimal collector tube outlet temperature trajectory under multiple constraints of "anti-freezing blockage - anti-overheating - reducing thermal stress". Summary of the Invention
[0007] The purpose of this invention is to provide a molten salt anti-condensation and safe start-up and shutdown control method based on generative adversarial networks, which solves multiple safety problems faced by molten salt in the collector tubes of linear Fresnel (LFR) solar thermal power plants during start-up and shutdown, such as freezing blockage, overheating and thermal stress damage. At the same time, it realizes the automatic generation of multi-region differentiated optimal temperature control trajectory.
[0008] The technical solution adopted in this invention is a molten salt anti-condensation and safe start-up / shutdown control method driven by generative adversarial networks, which is implemented according to the following steps: Step 1: Build a full-system multi-area heat collector pipe-molten salt coupled thermodynamic simulation model, including the main plant pipeline, mirror field branch heat collector pipe, mirror field main heat collector header, main plant-mirror field transmission pipeline section and valves; Step 2: Based on the auxiliary classifier condition generation mechanism, construct and train the multi-region, multi-constraint AC-GANs start trajectory generator offline; Step 3: Build and train the AC-GANs halt trajectory generator offline, and output differentiated heat preservation strategies; Step 4: Build a multi-regional coordinated model predictive controller; Step 5: Configure multi-level temperature hard protection logic; Step 6: Achieve complete closed-loop control.
[0009] The invention is further characterized in that, Step 1 is implemented in the following steps: To address the differences in thermal characteristics across different regions, heat balance equations were established for the main plant pipeline, the mirror field branch collector pipes, the mirror field main collector header, the transmission pipeline section between the main plant and the mirror field, and the valves. The mirror field branch adopts a fine-grained multi-node model, which includes an axial segmentation module, a radial layering module, a node energy balance module, and an adjacent node thermal coupling module. The axial segmentation module divides the heat collection tube into multiple calculation units along its length. The radial layering module sets up a molten salt main layer, a molten salt near-wall layer, an inner wall layer, a middle wall layer, and an outer wall layer within each axial calculation unit. The node energy balance module calculates the heat absorption, heat release, heat conduction, and convective heat transfer processes of each node layer. The adjacent node thermal coupling module is used to calculate the molten salt transport heat transfer and axial heat conduction of the tube wall between adjacent axial units, describing the molten salt convective heat transfer inside the tube, the axial heat conduction of the tube wall, and the solar radiation absorption process. The main path of the mirror field adopts a large-diameter thick-walled tube model. The large-diameter thick-walled tube model includes a thick-walled radial heat conduction module, a tube wall heat capacity module, a temperature-related physical property parameter module, and a temperature difference constraint module. The thick-walled radial heat conduction module divides the tube wall from the inside to the outside into an inner wall layer, a middle wall layer, and an outer wall layer, and calculates the heat transfer between each layer according to the cylindrical wall heat conduction equation. The tube wall heat capacity module is used to characterize the heat storage hysteresis of the thick-walled tube during the heating and cooling processes. The temperature-related physical property parameter module is used to correct the thermal conductivity, specific heat capacity, and density of the tube wall material at different temperatures. The temperature difference constraint module is used to calculate the temperature difference between the inner and outer walls and the maximum radial temperature difference of the tube wall, describing the temperature change response under large thermal inertia. For the main plant piping area, a low heat loss model is established considering the low heat dissipation environment indoors. The low heat loss model includes a solar absorption module, a selective absorption coating radiation module, a vacuum interlayer heat transfer module, a glass outer tube convection and radiation heat dissipation module, and a support component thermal bridge loss module. The solar absorption module calculates the effective solar radiation heat received by the collector tube; the selective absorption coating radiation module calculates the radiative heat transfer between the outer surface of the absorber tube and the glass outer tube; the vacuum interlayer heat transfer module describes the radiative heat transfer within the vacuum annulus; the glass outer tube convection and radiation heat dissipation module calculates the natural or forced convection heat dissipation and long-wave radiation heat dissipation between the glass outer tube and the environment; and the support component thermal bridge loss module calculates the additional heat loss caused by local structures such as supports and end connectors. A long-distance heat dissipation model is established for the pipeline section, incorporating ambient wind speed and thermal resistance of the insulation layer. The long-distance heat dissipation model includes a pipeline axial discretization module, a friction loss module, and an environmental boundary condition module. The pipeline axial discretization module divides the long-distance pipeline into multiple continuous pipe sections. The friction loss module calculates the heat released by each pipe section to the environment through the insulation layer. The environmental boundary condition module introduces ambient temperature, wind speed, and surface convective heat transfer coefficient to correct for heat dissipation intensity under different external conditions. A local heat transfer enhancement model for the narrowing and throttling flow in the valve area is established. This model includes an inlet contraction section module, a throat throttling section module, an outlet expansion section module, and an anti-condensation constraint module. The inlet contraction section module calculates the velocity change and local acceleration effect of molten salt entering the narrowing region. The throat throttling section module calculates the flow velocity, Reynolds number, and local heat transfer coefficient at the minimum cross-section of the narrowing. The outlet expansion section module calculates the flow recovery and local backflow effects after expansion. The anti-condensation constraint module determines whether the minimum temperature of molten salt in the narrowing region is lower than the solidification risk threshold and feeds this result back to the safe start-up and shutdown control process. Through joint simulation, key thermodynamic state quantities, including molten salt temperature, pipe wall temperature distribution and maximum temperature difference, inner wall temperature, outer wall temperature, and molten salt heating rate, are calculated in real time for each region.
[0010] In step 2, the auxiliary classifier uses five types of area labels—mirror field branch, mirror field main line, main plant, conveying pipeline section, and valve—as input conditions, forcing the generator to learn and distinguish the thermal characteristics of different areas. For the generation of the initial temperature rise trajectory, the mirror field branch uses the inlet molten salt temperature, the initial temperature field of the pipe wall, and the predicted solar irradiance as inputs, with the thermal stress safety threshold and the upper limit of the outlet temperature of 542℃ as hard constraints. Through the discrimination of the auxiliary classifier, the generator is forced to learn to use a step-by-step slow preheating in the low-temperature section to uniformly heat the pipe wall temperature, and actively slow down the heating rate when approaching the upper limit of operation, generating the optimal temperature rise curve for rapid response to irradiation. The conveying pipeline section is constrained by the minimum temperature along the line not being lower than the safety lower limit, generating a conveying trajectory that takes into account both anti-condensation and energy saving. The valve area is constrained by the maximum allowable temperature difference at the diameter reduction point, generating a gradually changing temperature trajectory to avoid thermal shock. During the offline training phase, a startup sample library was first constructed using a fine-grained multi-node model, a large-diameter thick-walled pipe model, a low heat loss model, a long-distance heat dissipation model, and a reduced-diameter throttling local heat transfer enhancement model. The sample library covers the startup process under different ambient temperatures, solar irradiance, cold salt tank outlet temperature, initial pipe wall temperature, molten salt flow rate, pipe length, pipe diameter, insulation heat loss, and reduced-diameter throttling structure parameters. First, the generator receives the startup condition vector:
[0011]
[0012] in, To start the initial molten salt temperature, The initial temperature of the inner wall of the pipe. The initial temperature of the outer wall of the pipe. For ambient temperature, Solar irradiance, For initial traffic, For the length of the pipeline, For pipe diameter, For heat loss per unit length, To ensure a safe lower limit, The maximum allowable temperature difference of the pipe wall, This is a random noise vector used to enable the generator to generate multiple sets of candidate trajectories under the same operating conditions. For parameters Generator network, The candidate trajectories output by the generator; The candidate trajectories output by the generator are represented as follows:
[0013] Set the trajectory for the candidate temperatures output by the generator; The candidate mass flow trajectory output by the generator, in units of; The candidate auxiliary heating power trajectory output by the generator; The candidate heating rate trajectory output by the generator; For time; After the generator outputs candidate trajectories, they enter the physical constraint layer, which modifies the candidate curves to obtain physically feasible trajectories.
[0014] The trajectory is corrected by the physical constraint layer; This is a physical constraint correction function; The candidate trajectories output by the generator; This is the current startup condition vector; The corrected trajectory is represented as follows:
[0015] To set the trajectory for the corrected temperature, The corrected mass flow trajectory. The corrected auxiliary heating power trajectory. This is the corrected heating rate trajectory; The discriminator distinguishes the corrected trajectory:
[0016] The safe startup probability or confidence score output by the discriminator has a value range of 0 to 1. For parameters The discriminator; The trajectory is corrected by the physical constraint layer; This is the current startup condition vector; The auxiliary classifier outputs the following operating condition categories:
[0017] This refers to the start-up operating condition category; For parameters Auxiliary classifier; The trajectory is corrected by the physical constraint layer; This is the current startup condition vector; During training, the generator parameters are adjusted in reverse based on the discriminant loss, classification loss, physical constraint loss, and smoothing loss. Total loss function:
[0018] Total loss function of the generator; : Adversarial loss, used to make the generated trajectory approximate the actual safe start trajectory distribution; : Auxiliary classification loss, used to ensure that the generated trajectory conforms to the corresponding startup condition category; Physical constraint loss is used to penalize unsafe conditions such as temperatures below the minimum, excessive heating rates, excessive pipe wall temperature differences, and excessive temperature drops along the pipe. Trajectory smoothing loss is used to reduce abrupt changes in temperature, flow rate, and heating power. : Classification loss weight coefficient; Physical constraint loss weighting coefficient; : Smoothing loss weighting coefficient; Physical constraint loss:
[0019] Physical constraint loss; Minimum temperature constraint penalty coefficient; : Pipe wall temperature difference constraint penalty coefficient; : Maximum heating rate constraint penalty coefficient; Minimum heating rate constraint penalty coefficient; The lowest permissible temperature for the entire system is 290℃. Temperature after physical constraint layer correction; Temperature difference between the inside and outside of the pipe wall; : Maximum allowable temperature difference of the pipe wall; : Heating rate after physical constraint layer correction; : Maximum permissible heating rate; Minimum allowable heating rate; Smoothing loss:
[0020] Trajectory smoothing loss; The corrected temperature at the current moment; Temperature corrected from the previous control cycle; Temperature after correction from the previous two control cycles; Adversarial loss function:
[0021] Adversarial loss on the generator side; The probability that the discriminator considers the corrected generated trajectory to be true. : Calculate the average loss value of all samples in the current batch; The auxiliary classification loss is in the form of the cross-entropy loss function:
[0022] : Auxiliary classification loss; Auxiliary classifier for trajectory determination Belongs to the real category The probability of; Enter the specified working condition category label, either in real or conditional form. : Input the trajectory of the auxiliary classifier.
[0023] In step 2, The startup temperature rise trajectory is a set of candidate trajectories output by the generator. The generator generates candidate startup temperature rise trajectories based on the initial molten salt temperature, pipe wall temperature, ambient temperature, solar irradiance, initial flow rate, pipe diameter, and pipe length. The physical constraint layer is based on the lower limit of molten salt anti-condensation safety, the maximum heating rate, and the maximum temperature difference of the pipe wall. Make corrections to obtain When the local molten salt temperature is between 290℃ and 310℃, the heating rate is 0.3℃ / min to 0.8℃ / min; when the local molten salt temperature is between 310℃ and 400℃, the heating rate can be increased to 0.8℃ / min to 1.5℃ / min; when the local molten salt temperature exceeds 400℃, the heating rate is controlled at 1.0℃ / min to 2.0℃ / min; the discriminator... Conduct a secure boot trustworthiness evaluation, and select those that meet security constraints and have high trustworthiness. As the output of the starting temperature rise trajectory; The optimal heating curve is the best trajectory selected from multiple candidate start-up trajectories output by the generator under the same operating conditions. Below, multiple sets of random noise vectors are input. This causes the generator to output multiple sets of candidate trajectories. Each candidate trajectory is corrected by a physical constraint layer to obtain the corresponding... The discriminator outputs the safe start probability or confidence score for each set of trajectories: A comprehensive evaluation was conducted, taking into account start-up time, temperature safety margin, pipe wall temperature difference, and the smoothness of the heating rate. Items with high safety and reliability, temperatures not lower than 290℃, and pipe wall temperature differences not exceeding [a certain threshold] were selected. The heating rate does not exceed the limit, and the changes in flow rate and heating power are relatively smooth. , as the optimal temperature rise curve; The delivery trajectory corresponds to the mass flow trajectory in the generator output. and the temperature trajectory related to pipe length, pipe diameter, and heat loss per unit length. After being corrected by the physical constraint layer, it forms and ; In the training sample library, the long-distance heat dissipation model will include the pipe length. Pipe diameter Heat loss per unit length Initial flow Ambient temperature Factors such as these are incorporated into the sample. After the generator is trained, it no longer calculates point-by-point according to a certain conveying temperature drop formula. Instead, it generates temperature settings, flow rates, and auxiliary heating trajectories that match the conveying process based on these input parameters. The auxiliary classifier outputs the operating condition category.
[0024] The auxiliary classifier uses five types of area labels: mirror field branch, mirror field main path, main plant, conveying pipeline section, and valve. When the trajectory corresponds to the conveying pipeline section area, This indicates the temperature setting trajectory of the delivery pipeline section. This represents the mass flow rate trajectory during the conveying process. This indicates the trajectory of auxiliary heating power required during the conveying process; The temperature trajectory after physical constraint layer correction corresponds to the gradually varying temperature trajectory. Furthermore, it limits its mutations through a smoothing loss, which is used during training:
[0025] The generator first outputs candidate temperature setpoint trajectories. The physical constraint layer corrects it to Ensure the temperature is not lower than 290℃, and that the heating rate and pipe wall temperature difference do not exceed limits. During training, [the following measures are taken]. By penalizing abrupt changes in the temperature trajectory, the generator gradually outputs a smoother temperature trajectory, ultimately yielding... This refers to a gradually changing temperature trajectory.
[0026] In step 3, The AC-GANs shutdown trajectory generator takes into account the predicted ambient temperature and shutdown duration, with a hard constraint of a minimum safe molten salt temperature of 250℃. For different areas: the mirror field branch and main mirror field generate matching cooling curves based on their respective heat dissipation characteristics and thermal inertia, allowing natural cooling within a safe range; the transport pipe section has high heat loss and freezing risk, so an intermittent heat tracing intervention trajectory is generated, allowing the molten salt inside the pipe to cool naturally within a safe range, with short-term heat replenishment intervention only initiated when the lower limit is reached; the valve area, as a weak point at the diameter reduction, generates a priority insulation trajectory, initiating heat tracing maintenance at the initial stage of shutdown to prevent preferential solidification at the diameter reduction point leading to blockage; the main plant pipeline area utilizes the relatively stable ambient temperature inside the plant to generate a passive insulation strategy, significantly reducing electric heat tracing energy consumption. Each area's trajectory is strictly penalized by a discriminator for results that cause the molten salt temperature inside the pipe to fall below the safe lower limit, while rewarding the generation result with the lowest total energy consumption, thereby minimizing plant power consumption during shutdown while ensuring no freezing. The shutdown trajectory generator uses a conditional deep convolutional generation structure, with the data flow from bottom to top, and consists of four components: (1) Input module: The input consists of three parts, namely the operating condition vector, the region category label, and the random noise vector. The operating condition vector represents the numerical vector of the operating conditions at the time of shutdown. The components include the initial temperature of molten salt, the initial temperature of the inner wall of the pipe, the initial temperature of the outer wall of the pipe, the ambient temperature, the initial mass flow rate, the pipe length, the pipe diameter, the heat loss per unit length, and the lower limit temperature of safety. Maximum allowable temperature difference of pipe wall The input is used as a deterministic condition for trajectory generation; the region category label is the region category of the pipe segment to which the trajectory to be generated belongs, taking one of five categories: mirror field branch, mirror field main road, main plant, conveying pipe segment, and valve; the label is encoded individually or embedded and mapped into a continuous label vector before participating in the fusion, serving as a condition constraint to distinguish the characteristics of trajectories in different regions; random noise vector. : Random vectors sampled according to a standard normal distribution. Under the same working condition vector and region category label, different noise samples correspond to different candidate trajectories, providing multiple sets of candidate samples for subsequent screening; (2) Input fusion node: After the working condition vector, label vector and random noise vector are aligned in dimension, they are fused into a unified initial feature tensor by adding element by element and then fed into the first convolutional basic unit. This node completes the injection of condition information and random information. (3) Convolutional feature basic unit stack: The fused features pass through n concatenated and repeated convolutional feature basic units in sequence. Each basic unit contains three layers in sequence: 3×3 convolutional layer: performs local convolution operation on time series features to extract local change features such as trend and curvature of trajectory; Normalization layer: Normalizes the convolution output, suppresses feature value drift, and stabilizes gradient propagation; ReLU activation layer: Outputs features with piecewise linear and nonlinear mapping, providing nonlinear expressive power; in The value is a positive integer, set according to the trajectory sampling length and feature dimension; multiple layers of basic units are stacked sequentially to form the feature extraction backbone of the generator; (4) Upsampling layer: Located after the stack of basic convolutional units, it upsamples the feature sequence along the time dimension, expands the low temporal resolution feature sequence to the set sampling length of the stopping trajectory, and completes the temporal reconstruction from coarse-scale features to full-length trajectory.
[0027] In step 3, (1) The matched cooling curve is used during the shutdown phase, and the generator outputs candidate temperature trajectories. But at this time The cooling rate trajectory is obtained after correction by the physical constraint layer. and ; Input the initial molten salt temperature at shutdown, initial inner wall temperature of the pipe, initial outer wall temperature of the pipe, ambient temperature, current flow rate, pipe length, pipe diameter, heat loss per unit length, safety lower limit, and maximum allowable temperature difference of the pipe wall into the operating condition vector. The generator outputs candidate stopping trajectories. ,in Indicates the candidate cooling curve. Indicates the candidate cooling rate, the physical constraint layer for Make corrections to ensure the corrected version The temperature should not be lower than 290℃, and the cooling rate, pipe wall temperature difference, and friction temperature drop should meet safety requirements. The discriminator should determine whether the cooling trajectory closely approximates the actual safe shutdown trajectory and meets the safe shutdown requirements. Output as a matching cooling curve; (2) The auxiliary heating power trajectory in the generator output corresponding to the intermittent heat tracing intervention trajectory After correction by the physical constraint layer, the following is obtained. ; Safety lower limit Initial molten salt temperature Ambient temperature Heat loss per unit length Pipeline length Pipe diameter Initial flow Maximum allowable temperature difference between pipe wall and pipe wall Input generator. Generator output: candidate auxiliary heating power trajectory. In the scenario of shutdown insulation, This indicates the change in heat tracing power over time. The physical constraint layer, based on the lower limit of temperature safety, the upper limit of heat tracing power, the rate of temperature change, and the smoothness requirements, [determines / imposes / adjusts] the [temperature] power. Revised to When the generator predicts the temperature trajectory Or the corrected temperature trajectory When the temperature approaches 290℃, the physical constraint layer increases the corresponding time period. When the temperature has a sufficient safety margin, reduce or shut off. This ultimately forms an intermittent heat tracing intervention trajectory of opening-maintaining-closing-reopening; (3) Priority insulation trajectory is used to determine the insulation priority of different areas at different times. The auxiliary classifier maps the trajectory to five types of area labels: mirror field branch, mirror field main road, main plant, conveying pipe section, and valve. For each type of area, the generator will generate the corresponding control trajectory based on its temperature change, heat loss, flow rate and auxiliary heating requirements. The generator outputs candidate trajectories under various operating conditions. The physical constraint layer was corrected to obtain Auxiliary classifier Identify the region category corresponding to the trajectory, and classify the regions accordingly. , , , , , and close The degree of heat preservation determines the priority order of regional insulation, with higher priority areas receiving more auxiliary heating power. Insulation time and flow rate maintenance strategy; (4) Passive insulation strategy is used to determine whether active heat tracing can be not started immediately after shutdown, and the system waste heat, insulation layer and low flow circulation can be used to maintain a safe temperature. Initial molten salt temperature Initial temperature of the inner wall of the pipe Initial temperature of the outer wall of the pipe Ambient temperature Pipeline length Pipe diameter Heat loss per unit length Initial flow and safety lower limit Input generator, generator outputs candidate temperature trajectories Candidate mass flow trajectory and candidate auxiliary heating power trajectory If the candidate trajectory output by the generator can still meet the requirements of a temperature not lower than 290℃, a pipe wall temperature difference not exceeding the limit, and a cooling rate not too fast when no or minimal auxiliary heating power is used, then the trajectory is determined to be a passive heat preservation candidate strategy, and the physical constraint layer modifies it to... The discriminator outputs a security and reliability score. .like If the value is greater than 0.75, then a passive insulation strategy will be output.
[0028] In step 4, Each of the five areas—mirror field branch, mirror field main road, main plant, conveying pipeline section, and valve—is equipped with a model predictive controller, resulting in a total of five model predictive controllers: mirror field branch model predictive controller, mirror field main road model predictive controller, main plant model predictive controller, conveying pipeline section model predictive controller, and valve model predictive controller. Each of these five controllers is independently responsible for temperature tracking and control within its respective area. The five model predictive controllers have the same structure, each consisting of five modules: (1) Input interface module: cycle The initial state is determined by collecting the measured values of the molten salt temperature and the inner / outer wall temperature of the pipe in this area at the beginning. Simultaneously, it receives the corrected trajectory output from the preceding "trajectory generation and physical constraint correction" step. Extract the prediction time domain Temperature components within As a future target trajectory; (2) Prediction model module: Stores the discrete state-space prediction model obtained by linearizing and discretizing the thermal model of this region through the working point, and outputs the predicted temperature sequence based on the initial state and candidate control sequence; (3) Rolling optimization module: Establishes and solves the constrained quadratic programming problem and outputs the optimal control sequence. Its first element is the core solution unit of the controller; (4) Control output module: The first element is decomposed into three set values: molten salt flow rate, electric heating power, and valve opening, which are sent to the flow regulating valve, the circulating pump frequency converter, the electric heater, and the electric actuator, respectively. (5) Coordination and communication module: Exchanges boundary temperature prediction sequences with adjacent area controllers and provides time-varying boundary conditions and inter-regional temperature difference coupling constraints to the rolling optimization module; In step 4, the model prediction controller operates as follows: (1) Status feedback and reference reception: period At the start time, the input interface module collects the measured values of the molten salt temperature and the inner / outer wall temperature of the pipe in this area. This constitutes the initial state. Simultaneously, it receives the corrected trajectory output from the preceding trajectory generation and physical constraint correction steps. Extract the prediction time domain The internal temperature component is used as the future target trajectory, denoted as... ; (2) Model prediction, time-domain preservation and feedback correction: The optimization variable is taken in the control time domain. Control sequence within , These are the molten salt flow rate, electric heating power, and valve opening, respectively, maintaining the following rule: when Time to take That is, control quantity is maintained Unchanged; accordingly The full-time control sequence is ,in To maintain the extended matrix: block rows ( Only in block columns Place Unit block The remaining blocks are all zero blocks, that is... The former Each block row forms a unit matrix, and each subsequent block row repeats the last block row, so that... Recursion for initial values:
[0029] Disturbance Constant within the time domain, its cumulative terms are incorporated. , to obtain the predicted temperature sequence , , Depend on Recursive construction, ensuring that the recursion is performed within the specified timeframe. Then proceed according to the rules; Feedback correction: Calculate prediction error Correct the predicted sequence , , Optimal choice 1; matrix form , Substitute the corrected sequence into the objective function and temperature constraints of step (3), and save the result in this cycle. For calculation in the next cycle This correction anchors the prediction to the measured value for each cycle, eliminating steady-state error. (3) Constrained rolling optimization and calculation of optimal control sequence: The objective function is
[0030] The tracked item is the corrected predicted value. The reference value is the future target trajectory received in step (1). ,Right now The two cannot be used interchangeably, and there are four categories of constraints: lower limit of temperature. Upper and lower limits of heating rate Actuator limit , Inter-regional boundary coupling constraints ; Calculation process: Let the difference matrix be set. The main diagonal of the block is The second diagonal of the block is , ,but , This represents the actual control quantity executed in the previous cycle; [record / denote] ( indivual), ,definition Substituting and expanding, we get the standard quadratic form. ,in
[0031]
[0032] All other constraints are Linear inequalities, written in a unified style Construction of the coupling constraint matrix: Let... To extract the boundary matrix, then Take the adjacent side boundary sequence Given a quantity, Expanding into a two-sided linear inequality, we get
[0033] Both lines are on the right. , It is an all-1 vector with no redundant constant terms; It is a fixed coefficient matrix that does not change with the coordination iteration, only Updated online with the swap sequence. All are constructed recursively from offline identification coefficients, requiring no online fitting. The rolling optimization module uses the effective set method or interior point method to solve the problem. ; Conflict constraint handling: hard constraints, lower temperature limit ≥ 290℃, actuator physical limit; soft constraints introduce slack variables. Relaxed to "original constraint" ",Add penalty to the objective function , If the value is a large positive number, and there is no solution, the constraints are relaxed in order of priority: first the coupling constraint, then the rate constraint, until a feasible solution is found; if there is still no solution, a safety fallback is triggered: the flow rate is taken as the upper limit of the anti-condensation cycle, and the heat tracing is taken as the rated power. After one cycle, the value is re-evaluated to ensure that there is a feasible solution in any cycle and the temperature is not lower than 290℃. (4) First element execution and coordination-driven: Only the first element is taken
[0034] The system is decomposed into three sets of parameters: the flow rate setpoint is sent to the flow regulating valve and the frequency converter of the circulating pump; the heat tracing power setpoint is sent to the thyristor power regulator; and the valve opening setpoint is sent to the electric actuator. These three types of actuators act on the thermal process of the heat collection pipes in this area, driving the actual temperature tracking. The remaining sequences are discarded, and the data is collected again in the next cycle and calculated according to step (2). Shift the time domain forward and then solve; (5) Inter-regional coordination, iteration, and convergence guarantee: Before distribution, communication is conducted with adjacent regions via a communication bus. Exchange the boundary temperature prediction sequences, and substitute the adjacent sequences as time-varying boundary conditions. and add coupling constraints .set up For the first The adjacent boundary sequence used in this round of solution is initially set to the latest adjacent exchange sequence. For the adjacent side The newly solved and returned boundary temperature prediction sequence is iterated in the order of solution-exchange-resolve, with three convergence mechanisms: (a) Damping exchange: This constitutes a compression mapping; Depend on Online calculation; (b) Termination condition: The maximum value of the difference between the two boundary boundaries is less than Or the number of iterations reaches That is, terminate; (c) Custody execution: to achieve If convergence is still not achieved, the latest solution element is executed, and residual mismatches are compensated by rolling feedback; The damping factor compresses the boundary mismatch by a fixed ratio of less than 1 in each round. The mismatch sequence is monotonically decreasing and has a lower bound of zero, so it will inevitably enter the threshold within a finite number of iterations. After the iteration terminates, the first element is executed to make the boundary temperature of adjacent regions consistent and avoid thermal stress.
[0035] Step 5 is implemented in the following steps: When the measured outlet temperature of the collector tube reaches the first-level warning threshold of 542℃, the collector power is forcibly reduced. The first-level warning reduces the collector power to 60%~80% of the rated collector power, and the reduction is 20%~40%. When the second-level alarm threshold of 545℃ is reached, the reduction is further increased. The second-level alarm reduces the collector power to 30%~50% of the rated collector power, that is, a cumulative reduction of 50%~70%. When the emergency state threshold of 547℃ is reached, heating is directly cut off and emergency cooling is initiated. When the temperature drops, a hysteresis range is set to prevent control oscillation. This hard protection logic ensures that even if the optimized trajectory generated by artificial intelligence deviates under extreme operating conditions, the actual operating state of the system will not exceed the equipment safety red line.
[0036] Step 6 is implemented in the following steps: The system first collects real-time data on molten salt temperature, pipe wall temperature, ambient temperature, solar irradiance, ambient wind speed, molten salt mass flow rate, and valve opening in each area using sensors. Then, the collected data is input into the AC-GANs trajectory generator, which generates target temperature trajectories for each area based on the current operating conditions and area category. The target trajectories are then passed to the model predictive controller (MMC) to calculate and execute the optimal control sequence—the third step in the MMC's operation. Simultaneously, multi-level temperature hard protection logic continuously monitors critical temperatures. This multi-level temperature hard protection logic includes the following protection strategies: when the measured collector tube outlet temperature reaches the first-level warning threshold... At 542℃, the heat collection power is rated to 70% of the rated heat collection power; when the secondary alarm threshold of 545℃ is reached, the heat collection power is rated to 40% of the rated heat collection power; when the emergency state threshold of 547℃ is reached, heating is directly cut off and emergency cooling is initiated, switching to a high-flow-rate cold salt circulation to rapidly cool the heat collection tubes; after each level of protection action, the heat collection power is allowed to be restored step by step only when the measured temperature falls back to the hysteresis range of 3℃ to 5℃ below the trigger threshold, to prevent control oscillation; under abnormal operating conditions, the protection action is triggered, and the entire control process forms a four-layer closed-loop control of data acquisition → trajectory generation → MPC control execution → safety protection.
[0037] The beneficial effects of this invention are: (1) This invention utilizes the multi-region condition generation capability of auxiliary classifier generative adversarial networks (AC-GANs) to automatically generate differentiated optimal temperature control trajectories for the differences in thermal inertia, heat dissipation conditions and safety margins of different regions in a linear Fresnel solar thermal power plant, breaking through the technical bottleneck of traditional fixed-parameter PID control or APS systems based on fixed rules that are difficult to dynamically optimize under multiple constraints. (2) This invention, through the condition generation mechanism of AC-GANs auxiliary classifiers, forces the generator to learn and distinguish the differences in thermal characteristics of five types of regions (main plant, mirror field branch, mirror field main road, conveying pipe section, valve), realizing differentiated trajectory generation with "one network for multiple uses", significantly reducing system complexity and training cost. (3) This invention uses a model predictive controller (MPC) as the lower-level execution mechanism, combined with independent multi-level temperature hard protection logic (542℃ / 545℃ / 547℃) as safety redundancy, to ensure that the optimized trajectory generated by AI will not exceed the equipment safety red line under extreme working conditions, thus ensuring the safety and reliability of the system. (4) By generating intermittent heat tracing intervention trajectory and passive heat preservation strategy during the shutdown process, this invention minimizes the plant power consumption during the shutdown period while ensuring that the molten salt in the pipe does not solidify, thus significantly improving the economic efficiency of power plant operation. Attached Figure Description
[0038] Figure 1 This is a structural diagram of AC-GANs. Detailed Implementation
[0039] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.
[0040] Figure 1 The diagram shows the structure of AC-GANs and illustrates the training process for generating the optimal temperature control trajectory.
[0041] The input vector is composed of operating conditions (inlet molten salt temperature of the collector tube, initial temperature field of the tube wall, predicted solar irradiance or ambient temperature, and downtime), regional category labels (five categories: mirror field branch, mirror field main, main plant, conveying pipe section, and valve), and random noise. The generator then upsamples layer by layer to generate candidate solutions for the original temperature trajectory. These candidate solutions are then entered into a physical constraint layer, where they are corrected based on hard constraints such as the tube wall thermal stress safety threshold, the upper limit of the outlet temperature (542℃), and the minimum safe temperature of the molten salt (250℃). Unreasonable sections that violate physical boundaries are eliminated, and the corrected temperature trajectory that satisfies multiple safety constraints is output. Subsequently, the corrected temperature trajectory and the actual temperature trajectory are input into a discriminator. After downsampling feature extraction, an auxiliary classifier determines the regional category of the input trajectory, while the discriminator determines whether the input trajectory is true or false. The classification loss and adversarial loss output by the discriminator are fed back to the generator and discriminator through backpropagation for alternating fine-tuning training. This allows the generator to gradually learn to generate more realistic temperature trajectories that match the thermal characteristics of each region, while continuously enhancing the discriminator's discrimination ability. The two work together in an adversarial game to iteratively optimize each other, ultimately achieving high-quality generation of multi-region differentiated optimal temperature control trajectories.
[0042] Part 1: System Architecture This invention discloses a molten salt anti-condensation and safe start-up / shutdown control method based on generative adversarial networks, targeting the binary nitrate working fluid in the collector tubes of a linear Fresnel (LFR) solar thermal power plant. Its system architecture consists of the following three core components: I. Multi-zone collector tube-molten salt coupled thermodynamic simulation model The model divides the molten salt piping system of the LFR solar thermal power plant into five regions: the main plant piping region, the mirror field branch collector pipe region, the mirror field main collector header region, the transmission pipe section between the main plant and the mirror field region, and the valve region. Based on the mechanistic model, segmented heat balance equations are established for each region, and key thermodynamic state quantities such as molten salt temperature, pipe wall temperature distribution, and maximum temperature difference are calculated in real time through joint simulation.
[0043] II. AC-GANs Trajectory Generator This component includes a multi-region start-up trajectory generator and a multi-region shutdown trajectory generator. An auxiliary classifier uses five region categories as condition labels, forcing the generator to learn the differences in thermal characteristics across different regions. The start-up trajectory generator takes the current inlet molten salt temperature and initial pipe wall temperature as input, and uses the pipe wall thermal stress safety threshold and the outlet temperature upper limit of 542℃ as hard constraints to generate the optimal target temperature rise curve for the molten salt at the collector pipe outlet. The shutdown trajectory generator takes the predicted ambient temperature and shutdown duration as input, and uses the minimum safe molten salt temperature of 250℃ as a hard constraint to generate the minimum insulation reference trajectory for the collector pipe outlet and inlet.
[0044] III. Model Predictive Controller (MPC) This component, acting as a lower-level actuator, receives the target temperature trajectory output in real time from the AC-GANs generator. Within each control cycle, based on the current measured temperature of each region and the predicted future target trajectory, it calculates the optimal control sequence by solving a constrained rolling optimization problem. By coordinating and adjusting the molten salt flow rate, electric heating power, and valve opening of each region, it drives the temperature of each region to accurately track its respective optimal reference curve.
[0045] Part Two: The Role of Each Component, Reasons for Selection, and Combination Principles The functions of each component in the method of this invention are as follows: Multi-zone collector tube-molten salt coupled thermodynamic simulation model: The model serves to provide physically realistic training data and constraints for the AC-GANs trajectory generator. Corresponding heat balance equations are established for each region based on its different thermal characteristics: a fine-grained multi-node model is used for the mirror field branch to describe the molten salt convection heat transfer inside the pipe, axial heat conduction through the pipe wall, and solar radiation absorption processes; a large-diameter, thick-walled pipe model is used for the mirror field main route to describe the temperature change response under high thermal inertia; a low-heat-loss model is established for the main plant pipeline area considering the low indoor heat dissipation environment; a long-distance heat dissipation model incorporating ambient wind speed and insulation thermal resistance is established for the transport pipe section area; and a local heat transfer enhancement model with reduced diameter and throttling is established for the valve area. The rationale for choosing this multi-regional differentiated modeling approach is that the thermal inertia, heat dissipation conditions, and safety margins differ significantly between different regions, and a single model cannot accurately describe the temperature change characteristics of each region.
[0046] AC-GANs trajectory generator: The function of this component is to automatically generate optimal temperature control trajectories for each region that satisfy multiple constraints (anti-freezing, anti-overheating, and thermal stress reduction). AC-GANs were chosen instead of ordinary GANs or conditional GANs because the auxiliary classifier mechanism of AC-GANs can utilize a single network to simultaneously learn the differences in thermal characteristics of five regions, achieving differentiated trajectory generation with "one network for multiple uses," significantly reducing system complexity and training costs. The startup trajectory generator employs a stepped, slow preheating strategy in the low-temperature range to uniformly heat the pipe wall, actively slowing the heating rate as it approaches the upper operating limit. The shutdown trajectory generator uses a discriminator to severely punish results that cause the molten salt temperature inside the pipe to fall below the safe lower limit, while rewarding the generation result with the lowest total energy consumption.
[0047] Model Predictive Controller (MPC): The function of this component is to act as a lower-level actuator, translating the target temperature trajectory generated by AC-GANs into specific control actions (molten salt flow rate, electric heating power, valve opening). The reason for choosing MPC instead of a traditional PID controller is that MPC can explicitly handle multivariable coupling constraints and achieve forward-looking control through rolling optimization, effectively addressing the temperature coordination problem between multiple regions.
[0048] The selection criteria for each component of this invention are as follows: This invention adopts a layered architecture of "upper-layer intelligent trajectory generation + lower-layer precise tracking control". The upper-layer AC-GANs trajectory generator, based on training data from a physical simulation model, utilizes the powerful nonlinear mapping capabilities of deep learning to automatically discover the optimal temperature control strategy that satisfies multiple constraints. The lower-layer MPC controller, on the other hand, leverages the predictive capabilities of the model to optimize the control sequence within a finite time domain, achieving precise tracking of the upper-layer trajectory. The upper and lower layers are coupled through the target temperature trajectory, while independent multi-level temperature hard protection logic (542℃ / 545℃ / 547℃) serves as safety redundancy, ensuring that the system will not exceed the equipment's safety limits under any operating conditions.
[0049] Part Three: Elements of Alternative Technical Features The following technical features in this invention have alternative equivalent implementations: (1) Replacement of trajectory generation network: AC-GANs can be replaced by other conditional generation models, such as conditional generative adversarial network (CGAN), information-maximizing generative adversarial network (InfoGAN) or conditional variational autoencoder (CVAE), as long as they can accept region category as conditional input and generate differentiated temperature trajectories.
[0050] (2) Replacement of lower-level controller: The model predictive controller (MPC) can be replaced by a fuzzy controller or a reinforcement learning controller, as long as it can receive the target temperature trajectory and drive accurate temperature tracking.
[0051] (3) Alternatives to the division of the five types of areas: The division of the five types of areas can be adjusted according to the actual power plant layout. For example, the mirror field branch can be further subdivided into sub-areas with different irradiation conditions, or multiple valve areas can be merged into a single area category.
[0052] (4) Replacement of temperature thresholds: The temperature protection thresholds at each level (542℃ / 545℃ / 547℃) and the minimum safe temperature (250℃) can be adjusted according to the specific molten salt working medium type and equipment specifications. The safety constraints of the start-up temperature rise trajectory and the safety constraints of the shutdown heat preservation trajectory can be reset according to the actual working conditions.
[0053] Part Four: Invention Point The core inventive point of this invention is: Invention Point 1: The auxiliary classifier mechanism of Generative Adversarial Networks (AC-GANs) is used to take five types of areas as classification inputs: main plant, mirror field branch, mirror field main road, conveying pipeline between main plant and mirror field, and valve. The optimal temperature control trajectory is generated according to the thermal inertia, heat dissipation conditions and safety margin of each area, which breaks through the technical bottleneck of traditional fixed parameter control that cannot adapt to the different chemical conditions of multiple areas.
[0054] Invention Point 2: Through the conditional generation mechanism of the AC-GANs auxiliary classifier, differentiated trajectory generation with "one network for multiple uses" is realized - a single network can generate the optimal temperature rise curve or heat preservation reference trajectory that matches the thermal characteristics of five types of regions, which significantly reduces system complexity and training cost.
[0055] Invention Point 3: The Model Predictive Controller (MPC) is used as the lower-level actuator. The controllers of each region coordinate by sharing boundary temperature information. The boundary temperature of adjacent regions is used as the coupling constraint of their respective MPC optimization problems to avoid thermal stress problems caused by excessive temperature difference between regions. At the same time, multi-level temperature hard protection logic (542℃ / 545℃ / 547℃) is independently configured as a safety redundancy mechanism to ensure that the optimization trajectory generated by AI will not exceed the equipment safety red line under extreme working conditions.
[0056] Invention Point 4: During shutdown, by generating intermittent heat tracing intervention trajectories and passive insulation strategies, differentiated insulation schemes are adopted for different areas—intermittent heat tracing intervention trajectories are generated for the delivery pipe section, priority insulation trajectories are generated for the valve area, and passive insulation strategies are generated for the main plant pipeline area—while ensuring that the molten salt inside the pipe does not solidify, the plant power consumption during shutdown is minimized.
[0057] Example 1 This invention is based on a generative adversarial network-driven method for molten salt anti-condensation and safe start-up / shutdown control, specifically implemented according to the following steps: Step 1: Construct a full-system multi-region thermodynamic simulation model of the coupling between the main plant pipeline, the mirror field branch collector pipes, the mirror field main collector header, the transmission pipeline section between the main plant and the mirror field, and the valves; Step 2: Construct and train an AC-GANs start trajectory generator with multiple regions and multiple constraints offline, making full use of the conditional generation mechanism of the auxiliary classifier; Step 3: Build and train the AC-GANs halt trajectory generator offline. Similarly, use the region category as an auxiliary classification condition to generate differentiated heat preservation strategies for different regions.
[0058] Step 4: Construct a multi-region coordinated model predictive controller (MPC) as a lower-level actuator to receive the target temperature trajectory output in real time by the AC-GANs generator. Step 5: Configure multi-level temperature hard protection logic independently outside the model predictive control loop as a safety redundancy mechanism.
[0059] Step 6: During the online operation phase, integrate the above modules into a complete closed-loop control system.
[0060] Example 2 This invention is based on a generative adversarial network-driven method for molten salt anti-condensation and safe start-up / shutdown control, specifically implemented according to the following steps: Step 1: Construct a full-system multi-region thermodynamic simulation model of the coupling between the main plant pipeline, the mirror field branch collector pipes, the mirror field main collector header, the transmission pipeline section between the main plant and the mirror field, and the valves; Step 1 is implemented in the following steps: To address the differences in thermal characteristics across different regions, heat balance equations were established for the main plant pipeline, the mirror field branch collector pipes, the mirror field main collector header, the transmission pipeline section between the main plant and the mirror field, and the valves. The mirror field branch adopts a fine-grained multi-node model, which includes an axial segmentation module, a radial layering module, a node energy balance module, and an adjacent node thermal coupling module. The axial segmentation module divides the heat collection tube into multiple calculation units along its length. The radial layering module sets up a molten salt main layer, a molten salt near-wall layer, an inner wall layer, a middle wall layer, and an outer wall layer within each axial calculation unit. The node energy balance module calculates the heat absorption, heat release, heat conduction, and convective heat transfer processes of each node layer. The adjacent node thermal coupling module is used to calculate the molten salt transport heat transfer and axial heat conduction of the tube wall between adjacent axial units, describing the molten salt convective heat transfer inside the tube, the axial heat conduction of the tube wall, and the solar radiation absorption process. The main path of the mirror field adopts a large-diameter thick-walled tube model. The large-diameter thick-walled tube model includes a thick-walled radial heat conduction module, a tube wall heat capacity module, a temperature-related physical property parameter module, and a temperature difference constraint module. The thick-walled radial heat conduction module divides the tube wall from the inside to the outside into an inner wall layer, a middle wall layer, and an outer wall layer, and calculates the heat transfer between each layer according to the cylindrical wall heat conduction equation. The tube wall heat capacity module is used to characterize the heat storage hysteresis of the thick-walled tube during the heating and cooling processes. The temperature-related physical property parameter module is used to correct the thermal conductivity, specific heat capacity, and density of the tube wall material at different temperatures. The temperature difference constraint module is used to calculate the temperature difference between the inner and outer walls and the maximum radial temperature difference of the tube wall, describing the temperature change response under large thermal inertia. For the main plant piping area, a low heat loss model is established considering the low heat dissipation environment indoors. The low heat loss model includes a solar absorption module, a selective absorption coating radiation module, a vacuum interlayer heat transfer module, a glass outer tube convection and radiation heat dissipation module, and a support component thermal bridge loss module. The solar absorption module calculates the effective solar radiation heat received by the collector tube; the selective absorption coating radiation module calculates the radiative heat transfer between the outer surface of the absorber tube and the glass outer tube; the vacuum interlayer heat transfer module describes the radiative heat transfer within the vacuum annulus; the glass outer tube convection and radiation heat dissipation module calculates the natural or forced convection heat dissipation and long-wave radiation heat dissipation between the glass outer tube and the environment; and the support component thermal bridge loss module calculates the additional heat loss caused by local structures such as supports and end connectors. A long-distance heat dissipation model is established for the pipeline section, incorporating ambient wind speed and thermal resistance of the insulation layer. The long-distance heat dissipation model includes a pipeline axial discretization module, a friction loss module, and an environmental boundary condition module. The pipeline axial discretization module divides the long-distance pipeline into multiple continuous pipe sections. The friction loss module calculates the heat released by each pipe section to the environment through the insulation layer. The environmental boundary condition module introduces ambient temperature, wind speed, and surface convective heat transfer coefficient to correct for heat dissipation intensity under different external conditions. A local heat transfer enhancement model for the narrowing and throttling flow in the valve area is established. This model includes an inlet contraction section module, a throat throttling section module, an outlet expansion section module, and an anti-condensation constraint module. The inlet contraction section module calculates the velocity change and local acceleration effect of molten salt entering the narrowing region. The throat throttling section module calculates the flow velocity, Reynolds number, and local heat transfer coefficient at the minimum cross-section of the narrowing. The outlet expansion section module calculates the flow recovery and local backflow effects after expansion. The anti-condensation constraint module determines whether the minimum temperature of molten salt in the narrowing region is lower than the solidification risk threshold and feeds this result back to the safe start-up and shutdown control process. Through joint simulation, key thermodynamic state quantities, including molten salt temperature, pipe wall temperature distribution and maximum temperature difference, inner wall temperature, outer wall temperature, and molten salt heating rate, are calculated in real time for each region.
[0061] Step 2: Construct and train an AC-GANs start trajectory generator with multiple regions and multiple constraints offline, making full use of the conditional generation mechanism of the auxiliary classifier; Step 3: Build and train the AC-GANs halt trajectory generator offline. Similarly, use the region category as an auxiliary classification condition to generate differentiated heat preservation strategies for different regions.
[0062] Step 4: Construct a multi-region coordinated model predictive controller (MPC) as a lower-level actuator to receive the target temperature trajectory output in real time by the AC-GANs generator. Step 5: Configure multi-level temperature hard protection logic independently outside the model predictive control loop as a safety redundancy mechanism.
[0063] Step 6: During the online operation phase, integrate the above modules into a complete closed-loop control system.
[0064] Example 3 This invention is based on a generative adversarial network-driven method for molten salt anti-condensation and safe start-up / shutdown control, specifically implemented according to the following steps: Step 1: Construct a full-system multi-region thermodynamic simulation model of the coupling between the main plant pipeline, the mirror field branch collector pipes, the mirror field main collector header, the transmission pipeline section between the main plant and the mirror field, and the valves; Step 1 is implemented in the following steps: To address the differences in thermal characteristics across different regions, heat balance equations were established for the main plant pipeline, the mirror field branch collector pipes, the mirror field main collector header, the transmission pipeline section between the main plant and the mirror field, and the valves. The mirror field branch adopts a fine-grained multi-node model, which includes an axial segmentation module, a radial layering module, a node energy balance module, and an adjacent node thermal coupling module. The axial segmentation module divides the heat collection tube into multiple calculation units along its length. The radial layering module sets up a molten salt main layer, a molten salt near-wall layer, an inner wall layer, a middle wall layer, and an outer wall layer within each axial calculation unit. The node energy balance module calculates the heat absorption, heat release, heat conduction, and convective heat transfer processes of each node layer. The adjacent node thermal coupling module is used to calculate the molten salt transport heat transfer and axial heat conduction of the tube wall between adjacent axial units, describing the molten salt convective heat transfer inside the tube, the axial heat conduction of the tube wall, and the solar radiation absorption process. The main path of the mirror field adopts a large-diameter thick-walled tube model. The large-diameter thick-walled tube model includes a thick-walled radial heat conduction module, a tube wall heat capacity module, a temperature-related physical property parameter module, and a temperature difference constraint module. The thick-walled radial heat conduction module divides the tube wall from the inside to the outside into an inner wall layer, a middle wall layer, and an outer wall layer, and calculates the heat transfer between each layer according to the cylindrical wall heat conduction equation. The tube wall heat capacity module is used to characterize the heat storage hysteresis of the thick-walled tube during the heating and cooling processes. The temperature-related physical property parameter module is used to correct the thermal conductivity, specific heat capacity, and density of the tube wall material at different temperatures. The temperature difference constraint module is used to calculate the temperature difference between the inner and outer walls and the maximum radial temperature difference of the tube wall, describing the temperature change response under large thermal inertia. For the main plant piping area, a low heat loss model is established considering the low heat dissipation environment indoors. The low heat loss model includes a solar absorption module, a selective absorption coating radiation module, a vacuum interlayer heat transfer module, a glass outer tube convection and radiation heat dissipation module, and a support component thermal bridge loss module. The solar absorption module calculates the effective solar radiation heat received by the collector tube; the selective absorption coating radiation module calculates the radiative heat transfer between the outer surface of the absorber tube and the glass outer tube; the vacuum interlayer heat transfer module describes the radiative heat transfer within the vacuum annulus; the glass outer tube convection and radiation heat dissipation module calculates the natural or forced convection heat dissipation and long-wave radiation heat dissipation between the glass outer tube and the environment; and the support component thermal bridge loss module calculates the additional heat loss caused by local structures such as supports and end connectors. A long-distance heat dissipation model is established for the pipeline section, incorporating ambient wind speed and thermal resistance of the insulation layer. The long-distance heat dissipation model includes a pipeline axial discretization module, a friction loss module, and an environmental boundary condition module. The pipeline axial discretization module divides the long-distance pipeline into multiple continuous pipe sections. The friction loss module calculates the heat released by each pipe section to the environment through the insulation layer. The environmental boundary condition module introduces ambient temperature, wind speed, and surface convective heat transfer coefficient to correct for heat dissipation intensity under different external conditions. A local heat transfer enhancement model for the narrowing and throttling flow in the valve area is established. This model includes an inlet contraction section module, a throat throttling section module, an outlet expansion section module, and an anti-condensation constraint module. The inlet contraction section module calculates the velocity change and local acceleration effect of molten salt entering the narrowing region. The throat throttling section module calculates the flow velocity, Reynolds number, and local heat transfer coefficient at the minimum cross-section of the narrowing. The outlet expansion section module calculates the flow recovery and local backflow effects after expansion. The anti-condensation constraint module determines whether the minimum temperature of molten salt in the narrowing region is lower than the solidification risk threshold and feeds this result back to the safe start-up and shutdown control process. Through joint simulation, key thermodynamic state quantities, including molten salt temperature, pipe wall temperature distribution and maximum temperature difference, inner wall temperature, outer wall temperature, and molten salt heating rate, are calculated in real time for each region.
[0065] Step 2: Construct and train an AC-GANs start trajectory generator with multiple regions and multiple constraints offline, making full use of the conditional generation mechanism of the auxiliary classifier; In step 2, the auxiliary classifier uses five types of area labels—mirror field branch, mirror field main line, main plant, conveying pipeline section, and valve—as input conditions, forcing the generator to learn and distinguish the thermal characteristics of different areas. For the generation of the initial temperature rise trajectory, the mirror field branch uses the inlet molten salt temperature, the initial temperature field of the pipe wall, and the predicted solar irradiance as inputs, with the thermal stress safety threshold and the upper limit of the outlet temperature of 542℃ as hard constraints. Through the discrimination of the auxiliary classifier, the generator is forced to learn to use a step-by-step slow preheating in the low-temperature section to uniformly heat the pipe wall temperature, and actively slow down the heating rate when approaching the upper limit of operation, generating the optimal temperature rise curve for rapid response to irradiation. The conveying pipeline section is constrained by the minimum temperature along the line not being lower than the safety lower limit, generating a conveying trajectory that takes into account both anti-condensation and energy saving. The valve area is constrained by the maximum allowable temperature difference at the diameter reduction point, generating a gradually changing temperature trajectory to avoid thermal shock. During the offline training phase, a startup sample library was first constructed using a fine-grained multi-node model, a large-diameter thick-walled pipe model, a low heat loss model, a long-distance heat dissipation model, and a reduced-diameter throttling local heat transfer enhancement model. The sample library covers the startup process under different ambient temperatures, solar irradiance, cold salt tank outlet temperature, initial pipe wall temperature, molten salt flow rate, pipe length, pipe diameter, insulation heat loss, and reduced-diameter throttling structure parameters. First, the generator receives the startup condition vector:
[0066]
[0067] in, To start the initial molten salt temperature, The initial temperature of the inner wall of the pipe. The initial temperature of the outer wall of the pipe. For ambient temperature, Solar irradiance, For initial traffic, For the length of the pipeline, For pipe diameter, For heat loss per unit length, For a safety lower limit, a temperature range of 285–295°C can be used, with 290°C being preferred. The maximum allowable temperature difference of the pipe wall, This is a random noise vector used to enable the generator to generate multiple sets of candidate trajectories under the same operating conditions. For parameters Generator network, The candidate trajectories output by the generator; The candidate trajectories output by the generator are represented as follows:
[0068] Set the trajectory for the candidate temperatures output by the generator; The candidate mass flow trajectory output by the generator, in units of; The candidate auxiliary heating power trajectory output by the generator; The candidate heating rate trajectory output by the generator; For time; After the generator outputs candidate trajectories, they enter the physical constraint layer, which modifies the candidate curves to obtain physically feasible trajectories.
[0069] The trajectory is corrected by the physical constraint layer; This is a physical constraint correction function; The candidate trajectories output by the generator; This is the current startup condition vector; The corrected trajectory is represented as follows:
[0070] To set the trajectory for the corrected temperature, The corrected mass flow trajectory. The corrected auxiliary heating power trajectory. This is the corrected heating rate trajectory; The discriminator distinguishes the corrected trajectory:
[0071] The safe startup probability or confidence score output by the discriminator has a value range of 0 to 1. For parameters The discriminator; The trajectory is corrected by the physical constraint layer; This is the current startup condition vector; The auxiliary classifier outputs the following operating condition categories:
[0072] This refers to the start-up operating condition category; For parameters Auxiliary classifier; The trajectory is corrected by the physical constraint layer; This is the current startup condition vector; During training, the generator parameters are adjusted in reverse based on the discriminant loss, classification loss, physical constraint loss, and smoothing loss. Total loss function:
[0073] Total loss function of the generator; : Adversarial loss, used to make the generated trajectory approximate the actual safe start trajectory distribution; : Auxiliary classification loss, used to ensure that the generated trajectory conforms to the corresponding startup condition category; Physical constraint loss is used to penalize unsafe conditions such as temperatures below the minimum, excessive heating rates, excessive pipe wall temperature differences, and excessive temperature drops along the pipe. Trajectory smoothing loss is used to reduce abrupt changes in temperature, flow rate, and heating power. : Classification loss weight coefficient; Physical constraint loss weighting coefficient; : Smoothing loss weighting coefficient; Physical constraint loss:
[0074] Physical constraint loss; Minimum temperature constraint penalty coefficient; : Pipe wall temperature difference constraint penalty coefficient; : Maximum heating rate constraint penalty coefficient; Minimum heating rate constraint penalty coefficient; The lowest permissible temperature for the entire system is 290℃. Temperature after physical constraint layer correction; Temperature difference between the inside and outside of the pipe wall; : Maximum allowable temperature difference of the pipe wall; : Heating rate after physical constraint layer correction; : Maximum permissible heating rate; Minimum allowable heating rate; Smoothing loss:
[0075] Trajectory smoothing loss; The corrected temperature at the current moment; Temperature corrected from the previous control cycle; Temperature after correction from the previous two control cycles; Adversarial loss function:
[0076] Adversarial loss on the generator side; The probability that the discriminator considers the corrected generated trajectory to be true. : Calculate the average loss value of all samples in the current batch; The auxiliary classification loss is in the form of the cross-entropy loss function:
[0077] : Auxiliary classification loss; Auxiliary classifier for trajectory determination Belongs to the real category The probability of; Enter the specified working condition category label, either in real or conditional form. : Input the trajectory of the auxiliary classifier.
[0078] In step 2, The startup temperature rise trajectory is a set of candidate trajectories output by the generator. The generator generates candidate startup temperature rise trajectories based on the initial molten salt temperature, pipe wall temperature, ambient temperature, solar irradiance, initial flow rate, pipe diameter, and pipe length. The physical constraint layer is based on the lower limit of molten salt anti-condensation safety, the maximum heating rate, and the maximum temperature difference of the pipe wall. Make corrections to obtain When the local molten salt temperature is between 290℃ and 310℃, the heating rate is 0.3℃ / min to 0.8℃ / min; when the local molten salt temperature is between 310℃ and 400℃, the heating rate can be increased to 0.8℃ / min to 1.5℃ / min; when the local molten salt temperature exceeds 400℃, the heating rate is controlled at 1.0℃ / min to 2.0℃ / min; the discriminator... Conduct a secure boot trustworthiness evaluation, and select those that meet security constraints and have high trustworthiness. As the output of the starting temperature rise trajectory; The optimal heating curve is the best trajectory selected from multiple candidate start-up trajectories output by the generator under the same operating conditions. Below, multiple sets of random noise vectors are input. This causes the generator to output multiple sets of candidate trajectories. Each candidate trajectory is corrected by a physical constraint layer to obtain the corresponding... The discriminator outputs the safe start probability or confidence score for each set of trajectories: A comprehensive evaluation was conducted, taking into account start-up time, temperature safety margin, pipe wall temperature difference, and the smoothness of the heating rate. Items with high safety and reliability, temperatures not lower than 290℃, and pipe wall temperature differences not exceeding [a certain threshold] were selected. The heating rate does not exceed the limit, and the changes in flow rate and heating power are relatively smooth. , as the optimal temperature rise curve; The delivery trajectory corresponds to the mass flow trajectory in the generator output. and the temperature trajectory related to pipe length, pipe diameter, and heat loss per unit length. After being corrected by the physical constraint layer, it forms and ; In the training sample library, the long-distance heat dissipation model will include the pipe length. Pipe diameter Heat loss per unit length Initial flow Ambient temperature Factors such as these are incorporated into the sample. After the generator is trained, it no longer calculates point-by-point according to a certain conveying temperature drop formula. Instead, it generates temperature settings, flow rates, and auxiliary heating trajectories that match the conveying process based on these input parameters. The auxiliary classifier outputs the operating condition category.
[0079] The auxiliary classifier uses five types of area labels: mirror field branch, mirror field main path, main plant, conveying pipeline section, and valve. When the trajectory corresponds to the conveying pipeline section area, This indicates the temperature setting trajectory of the delivery pipeline section. This represents the mass flow rate trajectory during the conveying process. This indicates the trajectory of auxiliary heating power required during the conveying process; The temperature trajectory after physical constraint layer correction corresponds to the gradually varying temperature trajectory. Furthermore, it limits its mutations through a smoothing loss, which is used during training:
[0080] The generator first outputs candidate temperature setpoint trajectories. The physical constraint layer corrects it to Ensure the temperature is not lower than 290℃, and that the heating rate and pipe wall temperature difference do not exceed limits. During training, [the following measures are taken]. By penalizing abrupt changes in the temperature trajectory, the generator gradually outputs a smoother temperature trajectory, ultimately yielding... This refers to a gradually changing temperature trajectory.
[0081] Step 3: Build and train the AC-GANs halt trajectory generator offline. Similarly, use the region category as an auxiliary classification condition to generate differentiated heat preservation strategies for different regions.
[0082] Step 4: Construct a multi-region coordinated model predictive controller (MPC) as a lower-level actuator to receive the target temperature trajectory output in real time by the AC-GANs generator. Step 5: Configure multi-level temperature hard protection logic independently outside the model predictive control loop as a safety redundancy mechanism.
[0083] Step 6: During the online operation phase, integrate the above modules into a complete closed-loop control system.
[0084] Example 4 This invention is based on a generative adversarial network-driven method for molten salt anti-condensation and safe start-up / shutdown control, specifically implemented according to the following steps: Step 1: Construct a full-system multi-region thermodynamic simulation model of the coupling between the main plant pipeline, the mirror field branch collector pipes, the mirror field main collector header, the transmission pipeline section between the main plant and the mirror field, and the valves; Step 1 is implemented in the following steps: To address the differences in thermal characteristics across different regions, heat balance equations were established for the main plant pipeline, the mirror field branch collector pipes, the mirror field main collector header, the transmission pipeline section between the main plant and the mirror field, and the valves. The mirror field branch adopts a fine-grained multi-node model, which includes an axial segmentation module, a radial layering module, a node energy balance module, and an adjacent node thermal coupling module. The axial segmentation module divides the heat collection tube into multiple calculation units along its length. The radial layering module sets up a molten salt main layer, a molten salt near-wall layer, an inner wall layer, a middle wall layer, and an outer wall layer within each axial calculation unit. The node energy balance module calculates the heat absorption, heat release, heat conduction, and convective heat transfer processes of each node layer. The adjacent node thermal coupling module is used to calculate the molten salt transport heat transfer and axial heat conduction of the tube wall between adjacent axial units, describing the molten salt convective heat transfer inside the tube, the axial heat conduction of the tube wall, and the solar radiation absorption process. The main path of the mirror field adopts a large-diameter thick-walled tube model. The large-diameter thick-walled tube model includes a thick-walled radial heat conduction module, a tube wall heat capacity module, a temperature-related physical property parameter module, and a temperature difference constraint module. The thick-walled radial heat conduction module divides the tube wall from the inside to the outside into an inner wall layer, a middle wall layer, and an outer wall layer, and calculates the heat transfer between each layer according to the cylindrical wall heat conduction equation. The tube wall heat capacity module is used to characterize the heat storage hysteresis of the thick-walled tube during the heating and cooling processes. The temperature-related physical property parameter module is used to correct the thermal conductivity, specific heat capacity, and density of the tube wall material at different temperatures. The temperature difference constraint module is used to calculate the temperature difference between the inner and outer walls and the maximum radial temperature difference of the tube wall, describing the temperature change response under large thermal inertia. For the main plant piping area, a low heat loss model is established considering the low heat dissipation environment indoors. The low heat loss model includes a solar absorption module, a selective absorption coating radiation module, a vacuum interlayer heat transfer module, a glass outer tube convection and radiation heat dissipation module, and a support component thermal bridge loss module. The solar absorption module calculates the effective solar radiation heat received by the collector tube; the selective absorption coating radiation module calculates the radiative heat transfer between the outer surface of the absorber tube and the glass outer tube; the vacuum interlayer heat transfer module describes the radiative heat transfer within the vacuum annulus; the glass outer tube convection and radiation heat dissipation module calculates the natural or forced convection heat dissipation and long-wave radiation heat dissipation between the glass outer tube and the environment; and the support component thermal bridge loss module calculates the additional heat loss caused by local structures such as supports and end connectors. A long-distance heat dissipation model is established for the pipeline section, incorporating ambient wind speed and thermal resistance of the insulation layer. The long-distance heat dissipation model includes a pipeline axial discretization module, a friction loss module, and an environmental boundary condition module. The pipeline axial discretization module divides the long-distance pipeline into multiple continuous pipe sections. The friction loss module calculates the heat released by each pipe section to the environment through the insulation layer. The environmental boundary condition module introduces ambient temperature, wind speed, and surface convective heat transfer coefficient to correct for heat dissipation intensity under different external conditions. A local heat transfer enhancement model for the narrowing and throttling flow in the valve area is established. This model includes an inlet contraction section module, a throat throttling section module, an outlet expansion section module, and an anti-condensation constraint module. The inlet contraction section module calculates the velocity change and local acceleration effect of molten salt entering the narrowing region. The throat throttling section module calculates the flow velocity, Reynolds number, and local heat transfer coefficient at the minimum cross-section of the narrowing. The outlet expansion section module calculates the flow recovery and local backflow effects after expansion. The anti-condensation constraint module determines whether the minimum temperature of molten salt in the narrowing region is lower than the solidification risk threshold and feeds this result back to the safe start-up and shutdown control process. Through joint simulation, key thermodynamic state quantities, including molten salt temperature, pipe wall temperature distribution and maximum temperature difference, inner wall temperature, outer wall temperature, and molten salt heating rate, are calculated in real time for each region.
[0085] Step 2: Construct and train an AC-GANs start trajectory generator with multiple regions and multiple constraints offline, making full use of the conditional generation mechanism of the auxiliary classifier; In step 2, the auxiliary classifier uses five types of area labels—mirror field branch, mirror field main line, main plant, conveying pipeline section, and valve—as input conditions, forcing the generator to learn and distinguish the thermal characteristics of different areas. For the generation of the initial temperature rise trajectory, the mirror field branch uses the inlet molten salt temperature, the initial temperature field of the pipe wall, and the predicted solar irradiance as inputs, with the thermal stress safety threshold and the upper limit of the outlet temperature of 542℃ as hard constraints. Through the discrimination of the auxiliary classifier, the generator is forced to learn to use a step-by-step slow preheating in the low-temperature section to uniformly heat the pipe wall temperature, and actively slow down the heating rate when approaching the upper limit of operation, generating the optimal temperature rise curve for rapid response to irradiation. The conveying pipeline section is constrained by the minimum temperature along the line not being lower than the safety lower limit, generating a conveying trajectory that takes into account both anti-condensation and energy saving. The valve area is constrained by the maximum allowable temperature difference at the diameter reduction point, generating a gradually changing temperature trajectory to avoid thermal shock. During the offline training phase, a startup sample library was first constructed using a fine-grained multi-node model, a large-diameter thick-walled pipe model, a low heat loss model, a long-distance heat dissipation model, and a reduced-diameter throttling local heat transfer enhancement model. The sample library covers the startup process under different ambient temperatures, solar irradiance, cold salt tank outlet temperature, initial pipe wall temperature, molten salt flow rate, pipe length, pipe diameter, insulation heat loss, and reduced-diameter throttling structure parameters. First, the generator receives the startup condition vector:
[0086]
[0087] in, To start the initial molten salt temperature, The initial temperature of the inner wall of the pipe. The initial temperature of the outer wall of the pipe. For ambient temperature, Solar irradiance, For initial traffic, For the length of the pipeline, For pipe diameter, For heat loss per unit length, For a safety lower limit, a temperature range of 285–295°C can be used, with 290°C being preferred. The maximum allowable temperature difference of the pipe wall, This is a random noise vector used to enable the generator to generate multiple sets of candidate trajectories under the same operating conditions. For parameters Generator network, The candidate trajectories output by the generator; The candidate trajectories output by the generator are represented as follows:
[0088] Set the trajectory for the candidate temperatures output by the generator; The candidate mass flow trajectory output by the generator, in units of; The candidate auxiliary heating power trajectory output by the generator; The candidate heating rate trajectory output by the generator; For time; After the generator outputs candidate trajectories, they enter the physical constraint layer, which modifies the candidate curves to obtain physically feasible trajectories.
[0089] The trajectory is corrected by the physical constraint layer; This is a physical constraint correction function; The candidate trajectories output by the generator; This is the current startup condition vector; The corrected trajectory is represented as follows:
[0090] To set the trajectory for the corrected temperature, The corrected mass flow trajectory. The corrected auxiliary heating power trajectory. This is the corrected heating rate trajectory; The discriminator distinguishes the corrected trajectory:
[0091] The safe startup probability or confidence score output by the discriminator has a value range of 0 to 1. For parameters The discriminator; The trajectory is corrected by the physical constraint layer; This is the current startup condition vector; The auxiliary classifier outputs the following operating condition categories:
[0092] This refers to the start-up operating condition category; For parameters Auxiliary classifier; The trajectory is corrected by the physical constraint layer; This is the current startup condition vector; During training, the generator parameters are adjusted in reverse based on the discriminant loss, classification loss, physical constraint loss, and smoothing loss. Total loss function:
[0093] Total loss function of the generator; : Adversarial loss, used to make the generated trajectory approximate the actual safe start trajectory distribution; : Auxiliary classification loss, used to ensure that the generated trajectory conforms to the corresponding startup condition category; Physical constraint loss is used to penalize unsafe conditions such as temperatures below the minimum, excessive heating rates, excessive pipe wall temperature differences, and excessive temperature drops along the pipe. Trajectory smoothing loss is used to reduce abrupt changes in temperature, flow rate, and heating power. : Classification loss weight coefficient; Physical constraint loss weighting coefficient; : Smoothing loss weighting coefficient; Physical constraint loss:
[0094] Physical constraint loss; Minimum temperature constraint penalty coefficient; : Pipe wall temperature difference constraint penalty coefficient; : Maximum heating rate constraint penalty coefficient; Minimum heating rate constraint penalty coefficient; The lowest permissible temperature for the entire system is 290℃. Temperature after physical constraint layer correction; Temperature difference between the inside and outside of the pipe wall; : Maximum allowable temperature difference of the pipe wall; : Heating rate after physical constraint layer correction; : Maximum permissible heating rate; Minimum allowable heating rate; Smoothing loss:
[0095] Trajectory smoothing loss; The corrected temperature at the current moment; Temperature corrected from the previous control cycle; Temperature after correction from the previous two control cycles; Adversarial loss function:
[0096] Adversarial loss on the generator side; The probability that the discriminator considers the corrected generated trajectory to be true. : Calculate the average loss value of all samples in the current batch; The auxiliary classification loss is in the form of the cross-entropy loss function:
[0097] : Auxiliary classification loss; Auxiliary classifier for trajectory determination Belongs to the real category The probability of; Enter the specified working condition category label, either in real or conditional form. : Input the trajectory of the auxiliary classifier.
[0098] In step 2, The startup temperature rise trajectory is a set of candidate trajectories output by the generator. The generator generates candidate startup temperature rise trajectories based on the initial molten salt temperature, pipe wall temperature, ambient temperature, solar irradiance, initial flow rate, pipe diameter, and pipe length. The physical constraint layer is based on the lower limit of molten salt anti-condensation safety, the maximum heating rate, and the maximum temperature difference of the pipe wall. Make corrections to obtain When the local molten salt temperature is between 290℃ and 310℃, the heating rate is 0.3℃ / min to 0.8℃ / min; when the local molten salt temperature is between 310℃ and 400℃, the heating rate can be increased to 0.8℃ / min to 1.5℃ / min; when the local molten salt temperature exceeds 400℃, the heating rate is controlled at 1.0℃ / min to 2.0℃ / min; the discriminator... Conduct a secure boot trustworthiness evaluation, and select those that meet security constraints and have high trustworthiness. As the output of the starting temperature rise trajectory; The optimal heating curve is the best trajectory selected from multiple candidate start-up trajectories output by the generator under the same operating conditions. Below, multiple sets of random noise vectors are input. This causes the generator to output multiple sets of candidate trajectories. Each candidate trajectory is corrected by a physical constraint layer to obtain the corresponding... The discriminator outputs the safe start probability or confidence score for each set of trajectories: A comprehensive evaluation was conducted, taking into account start-up time, temperature safety margin, pipe wall temperature difference, and the smoothness of the heating rate. Items with high safety and reliability, temperatures not lower than 290℃, and pipe wall temperature differences not exceeding [a certain threshold] were selected. The heating rate does not exceed the limit, and the changes in flow rate and heating power are relatively smooth. , as the optimal temperature rise curve; The delivery trajectory corresponds to the mass flow trajectory in the generator output. and the temperature trajectory related to pipe length, pipe diameter, and heat loss per unit length. After being corrected by the physical constraint layer, it forms and ; In the training sample library, the long-distance heat dissipation model will include the pipe length. Pipe diameter Heat loss per unit length Initial flow Ambient temperature Factors such as these are incorporated into the sample. After the generator is trained, it no longer calculates point-by-point according to a certain conveying temperature drop formula. Instead, it generates temperature settings, flow rates, and auxiliary heating trajectories that match the conveying process based on these input parameters. The auxiliary classifier outputs the operating condition category.
[0099] The auxiliary classifier uses five types of area labels: mirror field branch, mirror field main path, main plant, conveying pipeline section, and valve. When the trajectory corresponds to the conveying pipeline section area, This indicates the temperature setting trajectory of the delivery pipeline section. This represents the mass flow rate trajectory during the conveying process. This indicates the trajectory of auxiliary heating power required during the conveying process; The temperature trajectory after physical constraint layer correction corresponds to the gradually varying temperature trajectory. Furthermore, it limits its mutations through a smoothing loss, which is used during training:
[0100] The generator first outputs candidate temperature setpoint trajectories. The physical constraint layer corrects it to Ensure the temperature is not lower than 290℃, and that the heating rate and pipe wall temperature difference do not exceed limits. During training, [the following measures are taken]. By penalizing abrupt changes in the temperature trajectory, the generator gradually outputs a smoother temperature trajectory, ultimately yielding... This refers to a gradually changing temperature trajectory.
[0101] Step 3: Build and train the AC-GANs halt trajectory generator offline. Similarly, use the region category as an auxiliary classification condition to generate differentiated heat preservation strategies for different regions.
[0102] In step 3, The AC-GANs shutdown trajectory generator takes into account the predicted ambient temperature and shutdown duration, with a hard constraint of a minimum safe molten salt temperature of 250℃. For different areas: the mirror field branch and main mirror field generate matching cooling curves based on their respective heat dissipation characteristics and thermal inertia, allowing natural cooling within a safe range; the transport pipe section has high heat loss and freezing risk, so an intermittent heat tracing intervention trajectory is generated, allowing the molten salt inside the pipe to cool naturally within a safe range, with short-term heat replenishment intervention only initiated when the lower limit is reached; the valve area, as a weak point at the diameter reduction, generates a priority insulation trajectory, initiating heat tracing maintenance at the initial stage of shutdown to prevent preferential solidification at the diameter reduction point leading to blockage; the main plant pipeline area utilizes the relatively stable ambient temperature inside the plant to generate a passive insulation strategy, significantly reducing electric heat tracing energy consumption. Each area's trajectory is strictly penalized by a discriminator for results that cause the molten salt temperature inside the pipe to fall below the safe lower limit, while rewarding the generation result with the lowest total energy consumption, thereby minimizing plant power consumption during shutdown while ensuring no freezing. like Figure 1 As shown. The shutdown trajectory generator uses a conditional depthwise convolution generation structure, with the data flow from bottom to top, and consists of four components: (1) Input module: The input consists of three parts, namely the operating condition vector, the region category label, and the random noise vector. The operating condition vector represents the numerical vector of the operating conditions at the time of shutdown. The components include the initial temperature of molten salt, the initial temperature of the inner wall of the pipe, the initial temperature of the outer wall of the pipe, the ambient temperature, the initial mass flow rate, the pipe length, the pipe diameter, the heat loss per unit length, and the lower limit temperature of safety. Maximum allowable temperature difference of pipe wall The input is used as a deterministic condition for trajectory generation; the region category label is the region category of the pipe segment to which the trajectory to be generated belongs, taking one of five categories: mirror field branch, mirror field main road, main plant, conveying pipe segment, and valve; the label is encoded individually or embedded and mapped into a continuous label vector before participating in the fusion, serving as a condition constraint to distinguish the characteristics of trajectories in different regions; random noise vector. : Random vectors sampled according to a standard normal distribution. Under the same working condition vector and region category label, different noise samples correspond to different candidate trajectories, providing multiple sets of candidate samples for subsequent screening; (2) Input fusion node: After the working condition vector, label vector and random noise vector are aligned in dimension, they are fused into a unified initial feature tensor by adding element by element and then fed into the first convolutional basic unit. This node completes the injection of condition information and random information. (3) Convolutional feature basic unit stack: The fused features pass through n concatenated and repeated convolutional feature basic units in sequence. Each basic unit contains three layers in sequence: 3×3 convolutional layer (3×3Conv): performs local convolution operation on time series features to extract local change features such as trend and curvature of trajectory; Normalization layer (Norm): Normalizes the convolution output, suppresses feature value drift, and stabilizes gradient propagation; ReLU activation layer: Outputs features with piecewise linear and nonlinear mapping, providing nonlinear expressive power; in The value is a positive integer, set according to the trajectory sampling length and feature dimension; multiple layers of basic units are stacked sequentially to form the feature extraction backbone of the generator; (4) Upsampling layer: Located after the stack of basic convolutional units, it upsamples the feature sequence along the time dimension, expands the low temporal resolution feature sequence to the set sampling length of the stopping trajectory, and completes the temporal reconstruction from coarse-scale features to full-length trajectory.
[0103] In step 3, (1) The matched cooling curve is used during the shutdown phase, and the generator outputs candidate temperature trajectories. But at this time The cooling rate trajectory is obtained after correction by the physical constraint layer. and ; Input the initial molten salt temperature at shutdown, initial inner wall temperature of the pipe, initial outer wall temperature of the pipe, ambient temperature, current flow rate, pipe length, pipe diameter, heat loss per unit length, safety lower limit, and maximum allowable temperature difference of the pipe wall into the operating condition vector. The generator outputs candidate stopping trajectories. ,in Indicates the candidate cooling curve. Indicates the candidate cooling rate, the physical constraint layer for Make corrections to ensure the corrected version The temperature should not be lower than 290℃, and the cooling rate, pipe wall temperature difference, and friction temperature drop should meet safety requirements. The discriminator should determine whether the cooling trajectory closely approximates the actual safe shutdown trajectory and meets the safe shutdown requirements. Output as a matching cooling curve; (2) The auxiliary heating power trajectory in the generator output corresponding to the intermittent heat tracing intervention trajectory After correction by the physical constraint layer, the following is obtained. ; Safety lower limit Initial molten salt temperature Ambient temperature Heat loss per unit length Pipeline length Pipe diameter Initial flow Maximum allowable temperature difference between pipe wall and pipe wall Input generator. Generator output: candidate auxiliary heating power trajectory. In the scenario of shutdown insulation, This indicates the change in heat tracing power over time. The physical constraint layer, based on the lower limit of temperature safety, the upper limit of heat tracing power, the rate of temperature change, and the smoothness requirements, [determines / imposes / adjusts] the [temperature] power. Revised to When the generator predicts the temperature trajectory Or the corrected temperature trajectory When the temperature approaches 290℃, the physical constraint layer increases the corresponding time period. When the temperature has a sufficient safety margin, reduce or shut off. This ultimately forms an intermittent heat tracing intervention trajectory of opening-maintaining-closing-reopening; (3) Priority insulation trajectory is used to determine the insulation priority of different areas at different times. The auxiliary classifier maps the trajectory to five types of area labels: mirror field branch, mirror field main road, main plant, conveying pipeline section, and valve. For each type of area, the generator will generate a corresponding control trajectory based on its temperature change, heat loss, flow rate and auxiliary heating requirements. If a certain area Closer or the area Larger, pipeline length Longer, flow Lower, required If the value is higher, the area is considered a higher priority for insulation.
[0104] The generator outputs candidate trajectories under various operating conditions. The physical constraint layer was corrected to obtain Auxiliary classifier Identify the region category corresponding to the trajectory, and classify the regions accordingly. , , , , , and close The degree of heat preservation determines the priority order of regional insulation, with higher priority areas receiving more auxiliary heating power. Insulation time and flow rate maintenance strategy; (4) Passive insulation strategy is used to determine whether active heat tracing can be not started immediately after shutdown, and the system waste heat, insulation layer and low flow circulation can be used to maintain a safe temperature. Initial molten salt temperature Initial temperature of the inner wall of the pipe Initial temperature of the outer wall of the pipe Ambient temperature Pipeline length Pipe diameter Heat loss per unit length Initial flow and safety lower limit Input generator, generator outputs candidate temperature trajectories Candidate mass flow trajectory and candidate auxiliary heating power trajectory If the candidate trajectory output by the generator can still meet the requirements of a temperature not lower than 290℃, a pipe wall temperature difference not exceeding the limit, and a cooling rate not too fast when no or minimal auxiliary heating power is used, then the trajectory is determined to be a passive heat preservation candidate strategy, and the physical constraint layer modifies it to... The discriminator outputs a security and reliability score. .like If the value is greater than 0.75, then a passive insulation strategy will be output.
[0105] Step 4: Construct a multi-region coordinated model predictive controller (MPC) as a lower-level actuator to receive the target temperature trajectory output in real time by the AC-GANs generator. Step 5: Configure multi-level temperature hard protection logic independently outside the model predictive control loop as a safety redundancy mechanism.
[0106] Step 6: During the online operation phase, integrate the above modules into a complete closed-loop control system.
[0107] Example 5 This invention is based on a generative adversarial network-driven method for molten salt anti-condensation and safe start-up / shutdown control, specifically implemented according to the following steps: Step 1: Construct a full-system multi-region thermodynamic simulation model of the coupling between the main plant pipeline, the mirror field branch collector pipes, the mirror field main collector header, the transmission pipeline section between the main plant and the mirror field, and the valves; Step 1 is implemented in the following steps: To address the differences in thermal characteristics across different regions, heat balance equations were established for the main plant pipeline, the mirror field branch collector pipes, the mirror field main collector header, the transmission pipeline section between the main plant and the mirror field, and the valves. The mirror field branch adopts a fine-grained multi-node model, which includes an axial segmentation module, a radial layering module, a node energy balance module, and an adjacent node thermal coupling module. The axial segmentation module divides the heat collection tube into multiple calculation units along its length. The radial layering module sets up a molten salt main layer, a molten salt near-wall layer, an inner wall layer, a middle wall layer, and an outer wall layer within each axial calculation unit. The node energy balance module calculates the heat absorption, heat release, heat conduction, and convective heat transfer processes of each node layer. The adjacent node thermal coupling module is used to calculate the molten salt transport heat transfer and axial heat conduction of the tube wall between adjacent axial units, describing the molten salt convective heat transfer inside the tube, the axial heat conduction of the tube wall, and the solar radiation absorption process. The main path of the mirror field adopts a large-diameter thick-walled tube model. The large-diameter thick-walled tube model includes a thick-walled radial heat conduction module, a tube wall heat capacity module, a temperature-related physical property parameter module, and a temperature difference constraint module. The thick-walled radial heat conduction module divides the tube wall from the inside to the outside into an inner wall layer, a middle wall layer, and an outer wall layer, and calculates the heat transfer between each layer according to the cylindrical wall heat conduction equation. The tube wall heat capacity module is used to characterize the heat storage hysteresis of the thick-walled tube during the heating and cooling processes. The temperature-related physical property parameter module is used to correct the thermal conductivity, specific heat capacity, and density of the tube wall material at different temperatures. The temperature difference constraint module is used to calculate the temperature difference between the inner and outer walls and the maximum radial temperature difference of the tube wall, describing the temperature change response under large thermal inertia. For the main plant piping area, a low heat loss model is established considering the low heat dissipation environment indoors. The low heat loss model includes a solar absorption module, a selective absorption coating radiation module, a vacuum interlayer heat transfer module, a glass outer tube convection and radiation heat dissipation module, and a support component thermal bridge loss module. The solar absorption module calculates the effective solar radiation heat received by the collector tube; the selective absorption coating radiation module calculates the radiative heat transfer between the outer surface of the absorber tube and the glass outer tube; the vacuum interlayer heat transfer module describes the radiative heat transfer within the vacuum annulus; the glass outer tube convection and radiation heat dissipation module calculates the natural or forced convection heat dissipation and long-wave radiation heat dissipation between the glass outer tube and the environment; and the support component thermal bridge loss module calculates the additional heat loss caused by local structures such as supports and end connectors. A long-distance heat dissipation model is established for the pipeline section, incorporating ambient wind speed and thermal resistance of the insulation layer. The long-distance heat dissipation model includes a pipeline axial discretization module, a friction loss module, and an environmental boundary condition module. The pipeline axial discretization module divides the long-distance pipeline into multiple continuous pipe sections. The friction loss module calculates the heat released by each pipe section to the environment through the insulation layer. The environmental boundary condition module introduces ambient temperature, wind speed, and surface convective heat transfer coefficient to correct for heat dissipation intensity under different external conditions. A local heat transfer enhancement model for the narrowing and throttling flow in the valve area is established. This model includes an inlet contraction section module, a throat throttling section module, an outlet expansion section module, and an anti-condensation constraint module. The inlet contraction section module calculates the velocity change and local acceleration effect of molten salt entering the narrowing region. The throat throttling section module calculates the flow velocity, Reynolds number, and local heat transfer coefficient at the minimum cross-section of the narrowing. The outlet expansion section module calculates the flow recovery and local backflow effects after expansion. The anti-condensation constraint module determines whether the minimum temperature of molten salt in the narrowing region is lower than the solidification risk threshold and feeds this result back to the safe start-up and shutdown control process. Through joint simulation, key thermodynamic state quantities, including molten salt temperature, pipe wall temperature distribution and maximum temperature difference, inner wall temperature, outer wall temperature, and molten salt heating rate, are calculated in real time for each region.
[0108] Step 2: Construct and train an AC-GANs start trajectory generator with multiple regions and multiple constraints offline, making full use of the conditional generation mechanism of the auxiliary classifier; In step 2, the auxiliary classifier uses five types of area labels—mirror field branch, mirror field main line, main plant, conveying pipeline section, and valve—as input conditions, forcing the generator to learn and distinguish the thermal characteristics of different areas. For the generation of the initial temperature rise trajectory, the mirror field branch uses the inlet molten salt temperature, the initial temperature field of the pipe wall, and the predicted solar irradiance as inputs, with the thermal stress safety threshold and the upper limit of the outlet temperature of 542℃ as hard constraints. Through the discrimination of the auxiliary classifier, the generator is forced to learn to use a step-by-step slow preheating in the low-temperature section to uniformly heat the pipe wall temperature, and actively slow down the heating rate when approaching the upper limit of operation, generating the optimal temperature rise curve for rapid response to irradiation. The conveying pipeline section is constrained by the minimum temperature along the line not being lower than the safety lower limit, generating a conveying trajectory that takes into account both anti-condensation and energy saving. The valve area is constrained by the maximum allowable temperature difference at the diameter reduction point, generating a gradually changing temperature trajectory to avoid thermal shock. During the offline training phase, a startup sample library was first constructed using a fine-grained multi-node model, a large-diameter thick-walled pipe model, a low heat loss model, a long-distance heat dissipation model, and a reduced-diameter throttling local heat transfer enhancement model. The sample library covers the startup process under different ambient temperatures, solar irradiance, cold salt tank outlet temperature, initial pipe wall temperature, molten salt flow rate, pipe length, pipe diameter, insulation heat loss, and reduced-diameter throttling structure parameters. First, the generator receives the startup condition vector:
[0109]
[0110] in, To start the initial molten salt temperature, The initial temperature of the inner wall of the pipe. The initial temperature of the outer wall of the pipe. For ambient temperature, Solar irradiance, For initial traffic, For the length of the pipeline, For pipe diameter, For heat loss per unit length, For a safety lower limit, a temperature range of 285–295°C can be used, with 290°C being preferred. The maximum allowable temperature difference of the pipe wall, This is a random noise vector used to enable the generator to generate multiple sets of candidate trajectories under the same operating conditions. For parameters Generator network, The candidate trajectories output by the generator; The candidate trajectories output by the generator are represented as follows:
[0111] Set the trajectory for the candidate temperatures output by the generator; The candidate mass flow trajectory output by the generator, in units of; The candidate auxiliary heating power trajectory output by the generator; The candidate heating rate trajectory output by the generator; For time; After the generator outputs candidate trajectories, they enter the physical constraint layer, which modifies the candidate curves to obtain physically feasible trajectories.
[0112] The trajectory is corrected by the physical constraint layer; This is a physical constraint correction function; The candidate trajectories output by the generator; This is the current startup condition vector; The corrected trajectory is represented as follows:
[0113] To set the trajectory for the corrected temperature, The corrected mass flow trajectory. The corrected auxiliary heating power trajectory. This is the corrected heating rate trajectory; The discriminator distinguishes the corrected trajectory:
[0114] The safe startup probability or confidence score output by the discriminator has a value range of 0 to 1. For parameters The discriminator; The trajectory is corrected by the physical constraint layer; This is the current startup condition vector; The auxiliary classifier outputs the following operating condition categories:
[0115] This refers to the start-up operating condition category; For parameters Auxiliary classifier; The trajectory is corrected by the physical constraint layer; This is the current startup condition vector; During training, the generator parameters are adjusted in reverse based on the discriminant loss, classification loss, physical constraint loss, and smoothing loss. Total loss function:
[0116] Total loss function of the generator; : Adversarial loss, used to make the generated trajectory approximate the actual safe start trajectory distribution; : Auxiliary classification loss, used to ensure that the generated trajectory conforms to the corresponding startup condition category; Physical constraint loss is used to penalize unsafe conditions such as temperatures below the minimum, excessive heating rates, excessive pipe wall temperature differences, and excessive temperature drops along the pipe. Trajectory smoothing loss is used to reduce abrupt changes in temperature, flow rate, and heating power. : Classification loss weight coefficient; Physical constraint loss weighting coefficient; : Smoothing loss weighting coefficient; Physical constraint loss:
[0117] Physical constraint loss; Minimum temperature constraint penalty coefficient; : Pipe wall temperature difference constraint penalty coefficient; : Maximum heating rate constraint penalty coefficient; Minimum heating rate constraint penalty coefficient; The lowest permissible temperature for the entire system is 290℃. Temperature after physical constraint layer correction; Temperature difference between the inside and outside of the pipe wall; : Maximum allowable temperature difference of the pipe wall; : Heating rate after physical constraint layer correction; : Maximum permissible heating rate; Minimum allowable heating rate; Smoothing loss:
[0118] Trajectory smoothing loss; The corrected temperature at the current moment; Temperature corrected from the previous control cycle; Temperature after correction from the previous two control cycles; Adversarial loss function:
[0119] Adversarial loss on the generator side; The probability that the discriminator considers the corrected generated trajectory to be true. : Calculate the average loss value of all samples in the current batch; The auxiliary classification loss is in the form of the cross-entropy loss function:
[0120] : Auxiliary classification loss; Auxiliary classifier for trajectory determination Belongs to the real category The probability of; Enter the specified working condition category label, either in real or conditional form. : Input the trajectory of the auxiliary classifier.
[0121] In step 2, The startup temperature rise trajectory is a set of candidate trajectories output by the generator. The generator generates candidate startup temperature rise trajectories based on the initial molten salt temperature, pipe wall temperature, ambient temperature, solar irradiance, initial flow rate, pipe diameter, and pipe length. The physical constraint layer is based on the lower limit of molten salt anti-condensation safety, the maximum heating rate, and the maximum temperature difference of the pipe wall. Make corrections to obtain When the local molten salt temperature is between 290℃ and 310℃, the heating rate is 0.3℃ / min to 0.8℃ / min; when the local molten salt temperature is between 310℃ and 400℃, the heating rate can be increased to 0.8℃ / min to 1.5℃ / min; when the local molten salt temperature exceeds 400℃, the heating rate is controlled at 1.0℃ / min to 2.0℃ / min; the discriminator... Conduct a secure boot trustworthiness evaluation, and select those that meet security constraints and have high trustworthiness. As the output of the starting temperature rise trajectory; The optimal heating curve is the best trajectory selected from multiple candidate start-up trajectories output by the generator under the same operating conditions. Below, multiple sets of random noise vectors are input. This causes the generator to output multiple sets of candidate trajectories. Each candidate trajectory is corrected by a physical constraint layer to obtain the corresponding... The discriminator outputs the safe start probability or confidence score for each set of trajectories: A comprehensive evaluation was conducted, taking into account start-up time, temperature safety margin, pipe wall temperature difference, and the smoothness of the heating rate. Items with high safety and reliability, temperatures not lower than 290℃, and pipe wall temperature differences not exceeding [a certain threshold] were selected. The heating rate does not exceed the limit, and the changes in flow rate and heating power are relatively smooth. , as the optimal temperature rise curve; The delivery trajectory corresponds to the mass flow trajectory in the generator output. and the temperature trajectory related to pipe length, pipe diameter, and heat loss per unit length. After being corrected by the physical constraint layer, it forms and ; In the training sample library, the long-distance heat dissipation model will include the pipe length. Pipe diameter Heat loss per unit length Initial flow Ambient temperature Factors such as these are incorporated into the sample. After the generator is trained, it no longer calculates point-by-point according to a certain conveying temperature drop formula. Instead, it generates temperature settings, flow rates, and auxiliary heating trajectories that match the conveying process based on these input parameters. The auxiliary classifier outputs the operating condition category.
[0122] The auxiliary classifier uses five types of area labels: mirror field branch, mirror field main path, main plant, conveying pipeline section, and valve. When the trajectory corresponds to the conveying pipeline section area, This indicates the temperature setting trajectory of the delivery pipeline section. This represents the mass flow rate trajectory during the conveying process. This indicates the trajectory of auxiliary heating power required during the conveying process; The temperature trajectory after physical constraint layer correction corresponds to the gradually varying temperature trajectory. Furthermore, it limits its mutations through a smoothing loss, which is used during training:
[0123] The generator first outputs candidate temperature setpoint trajectories. The physical constraint layer corrects it to Ensure the temperature is not lower than 290℃, and that the heating rate and pipe wall temperature difference do not exceed limits. During training, [the following measures are taken]. By penalizing abrupt changes in the temperature trajectory, the generator gradually outputs a smoother temperature trajectory, ultimately yielding... This refers to a gradually changing temperature trajectory.
[0124] Step 3: Build and train the AC-GANs halt trajectory generator offline. Similarly, use the region category as an auxiliary classification condition to generate differentiated heat preservation strategies for different regions.
[0125] In step 3, The AC-GANs shutdown trajectory generator takes into account the predicted ambient temperature and shutdown duration, with a hard constraint of a minimum safe molten salt temperature of 250℃. For different areas: the mirror field branch and main mirror field generate matching cooling curves based on their respective heat dissipation characteristics and thermal inertia, allowing natural cooling within a safe range; the transport pipe section has high heat loss and freezing risk, so an intermittent heat tracing intervention trajectory is generated, allowing the molten salt inside the pipe to cool naturally within a safe range, with short-term heat replenishment intervention only initiated when the lower limit is reached; the valve area, as a weak point at the diameter reduction, generates a priority insulation trajectory, initiating heat tracing maintenance at the initial stage of shutdown to prevent preferential solidification at the diameter reduction point leading to blockage; the main plant pipeline area utilizes the relatively stable ambient temperature inside the plant to generate a passive insulation strategy, significantly reducing electric heat tracing energy consumption. Each area's trajectory is strictly penalized by a discriminator for results that cause the molten salt temperature inside the pipe to fall below the safe lower limit, while rewarding the generation result with the lowest total energy consumption, thereby minimizing plant power consumption during shutdown while ensuring no freezing. like Figure 1 As shown. The shutdown trajectory generator uses a conditional depthwise convolution generation structure, with the data flow from bottom to top, and consists of four components: (1) Input module: The input consists of three parts, namely the operating condition vector, the region category label, and the random noise vector. The operating condition vector represents the numerical vector of the operating conditions at the time of shutdown. The components include the initial temperature of molten salt, the initial temperature of the inner wall of the pipe, the initial temperature of the outer wall of the pipe, the ambient temperature, the initial mass flow rate, the pipe length, the pipe diameter, the heat loss per unit length, and the lower limit temperature of safety. Maximum allowable temperature difference of pipe wall The input is used as a deterministic condition for trajectory generation; the region category label is the region category of the pipe segment to which the trajectory to be generated belongs, taking one of five categories: mirror field branch, mirror field main road, main plant, conveying pipe segment, and valve; the label is encoded individually or embedded and mapped into a continuous label vector before participating in the fusion, serving as a condition constraint to distinguish the characteristics of trajectories in different regions; random noise vector. : Random vectors sampled according to a standard normal distribution. Under the same working condition vector and region category label, different noise samples correspond to different candidate trajectories, providing multiple sets of candidate samples for subsequent screening; (2) Input fusion node: After the working condition vector, label vector and random noise vector are aligned in dimension, they are fused into a unified initial feature tensor by adding element by element and then fed into the first convolutional basic unit. This node completes the injection of condition information and random information. (3) Convolutional feature basic unit stack: The fused features pass through n concatenated and repeated convolutional feature basic units in sequence. Each basic unit contains three layers in sequence: 3×3 convolutional layer (3×3Conv): performs local convolution operation on time series features to extract local change features such as trend and curvature of trajectory; Normalization layer (Norm): Normalizes the convolution output, suppresses feature value drift, and stabilizes gradient propagation; ReLU activation layer: Outputs features with piecewise linear and nonlinear mapping, providing nonlinear expressive power; in The value is a positive integer, set according to the trajectory sampling length and feature dimension; multiple layers of basic units are stacked sequentially to form the feature extraction backbone of the generator; (4) Upsampling layer: Located after the stack of basic convolutional units, it upsamples the feature sequence along the time dimension, expands the low temporal resolution feature sequence to the set sampling length of the stopping trajectory, and completes the temporal reconstruction from coarse-scale features to full-length trajectory.
[0126] In step 3, (1) The matched cooling curve is used during the shutdown phase, and the generator outputs candidate temperature trajectories. But at this time The cooling rate trajectory is obtained after correction by the physical constraint layer. and ; Input the initial molten salt temperature at shutdown, initial inner wall temperature of the pipe, initial outer wall temperature of the pipe, ambient temperature, current flow rate, pipe length, pipe diameter, heat loss per unit length, safety lower limit, and maximum allowable temperature difference of the pipe wall into the operating condition vector. The generator outputs candidate stopping trajectories. ,in Indicates the candidate cooling curve. Indicates the candidate cooling rate, the physical constraint layer for Make corrections to ensure the corrected version The temperature should not be lower than 290℃, and the cooling rate, pipe wall temperature difference, and friction temperature drop should meet safety requirements. The discriminator should determine whether the cooling trajectory closely approximates the actual safe shutdown trajectory and meets the safe shutdown requirements. Output as a matching cooling curve; (2) The auxiliary heating power trajectory in the generator output corresponding to the intermittent heat tracing intervention trajectory After correction by the physical constraint layer, the following is obtained. ; Safety lower limit Initial molten salt temperature Ambient temperature Heat loss per unit length Pipeline length Pipe diameter Initial flow Maximum allowable temperature difference between pipe wall and pipe wall Input generator. Generator output: candidate auxiliary heating power trajectory. In the scenario of shutdown insulation, This indicates the change in heat tracing power over time. The physical constraint layer, based on the lower limit of temperature safety, the upper limit of heat tracing power, the rate of temperature change, and the smoothness requirements, [determines / imposes / adjusts] the [temperature] power. Revised to When the generator predicts the temperature trajectory Or the corrected temperature trajectory When the temperature approaches 290℃, the physical constraint layer increases the corresponding time period. When the temperature has a sufficient safety margin, reduce or shut off. This ultimately forms an intermittent heat tracing intervention trajectory of opening-maintaining-closing-reopening; (3) Priority insulation trajectory is used to determine the insulation priority of different areas at different times. The auxiliary classifier maps the trajectory to five types of area labels: mirror field branch, mirror field main road, main plant, conveying pipeline section, and valve. For each type of area, the generator will generate a corresponding control trajectory based on its temperature change, heat loss, flow rate and auxiliary heating requirements. If a certain area Closer or the area Larger, pipeline length Longer, flow Lower, required If the value is higher, the area is considered a higher priority for insulation.
[0127] The generator outputs candidate trajectories under various operating conditions. The physical constraint layer was corrected to obtain Auxiliary classifier Identify the region category corresponding to the trajectory, and classify the regions accordingly. , , , , , and close The degree of heat preservation determines the priority order of regional insulation, with higher priority areas receiving more auxiliary heating power. Insulation time and flow rate maintenance strategy; (4) Passive insulation strategy is used to determine whether active heat tracing can be not started immediately after shutdown, and the system waste heat, insulation layer and low flow circulation can be used to maintain a safe temperature. Initial molten salt temperature Initial temperature of the inner wall of the pipe Initial temperature of the outer wall of the pipe Ambient temperature Pipeline length Pipe diameter Heat loss per unit length Initial flow and safety lower limit Input generator, generator outputs candidate temperature trajectories Candidate mass flow trajectory and candidate auxiliary heating power trajectory If the candidate trajectory output by the generator can still meet the requirements of a temperature not lower than 290℃, a pipe wall temperature difference not exceeding the limit, and a cooling rate not too fast when no or minimal auxiliary heating power is used, then the trajectory is determined to be a passive heat preservation candidate strategy, and the physical constraint layer modifies it to... The discriminator outputs a security and reliability score. .like If the value is greater than 0.75, then a passive insulation strategy will be output.
[0128] Step 4: Construct a multi-region coordinated model predictive controller (MPC) as a lower-level actuator to receive the target temperature trajectory output in real time by the AC-GANs generator. In step 4, Each of the five areas—mirror field branch, mirror field main road, main plant, conveying pipeline section, and valve—is equipped with a Model Predictive Controller (MPC), for a total of five MPCs: mirror field branch model predictive controller, mirror field main road model predictive controller, main plant model predictive controller, conveying pipeline section model predictive controller, and valve model predictive controller. Each of these five controllers is independently responsible for temperature tracking and control within its respective area. The five model predictive controllers have the same structure, each consisting of five modules: (1) Input interface module: cycle The initial state is determined by collecting the measured values of the molten salt temperature and the inner / outer wall temperature of the pipe in this area at the beginning. Simultaneously, it receives the corrected trajectory output from the preceding "trajectory generation and physical constraint correction" step. Extract the prediction time domain Temperature components within As a future target trajectory; (2) Prediction model module: Stores the discrete state-space prediction model obtained by linearizing and discretizing the thermal model of this region through the working point, and outputs the predicted temperature sequence based on the initial state and candidate control sequence; (3) Rolling optimization module: Establishes and solves the constrained quadratic programming problem and outputs the optimal control sequence. Its first element is the core solution unit of the controller; (4) Control output module: The first element is decomposed into three set values: molten salt flow rate, electric heating power, and valve opening, which are sent to the flow regulating valve, the circulating pump frequency converter, the electric heater, and the electric actuator, respectively. (5) Coordination and communication module: Exchanges boundary temperature prediction sequences with adjacent area controllers and provides time-varying boundary conditions and inter-regional temperature difference coupling constraints to the rolling optimization module; In step 4, the model prediction controller operates as follows: (1) Status feedback and reference reception: period At the start time, the input interface module collects the measured values of the molten salt temperature and the inner / outer wall temperature of the pipe in this area. This constitutes the initial state. Simultaneously, it receives the corrected trajectory output from the preceding trajectory generation and physical constraint correction steps. Extract the prediction time domain The internal temperature component is used as the future target trajectory, denoted as... ; (2) Model prediction, time-domain preservation and feedback correction: The optimization variable is taken in the control time domain. Control sequence within , These are the molten salt flow rate, electric heating power, and valve opening, respectively, maintaining the following rule: when Time to take That is, control quantity is maintained Unchanged; accordingly The full-time control sequence is ,in To maintain the extended matrix: block rows ( Only in block columns Place Unit block The remaining blocks are all zero blocks, that is... The former Each block row forms a unit matrix, and each subsequent block row repeats the last block row, so that... Recursion for initial values:
[0129] (disturbance) Constant within the time domain, its cumulative terms are incorporated. ), to obtain the predicted temperature sequence , , Depend on Recursive construction, ensuring that the recursion is performed within the specified timeframe. Then proceed according to the rules; Feedback correction: Calculate prediction error (The difference between the measured value and the predicted value for the current moment in the previous period), correct the prediction sequence. , , Optimal choice 1; matrix form , Substitute the corrected sequence into the objective function and temperature constraints of step (3), and save the result in this cycle. For calculation in the next cycle This correction anchors the prediction to the measured value for each cycle, eliminating steady-state error. (3) Constrained rolling optimization and calculation of optimal control sequence: The objective function is
[0130] The tracked item is the corrected predicted value. The reference value is the future target trajectory received in step (1). ,Right now The two cannot be used interchangeably, and there are four categories of constraints: lower limit of temperature. Upper and lower limits of heating rate Actuator limit , Inter-regional boundary coupling constraints (The adjacent boundary sequence is provided by step (5)); Calculation process: Set up the difference (displacement) matrix The main diagonal of the block is The second diagonal of the block is , ,but , This represents the actual control quantity executed in the previous cycle; [record / denote] ( indivual), ( (one), definition Substituting and expanding, we get the standard quadratic form. ,in
[0131]
[0132] All other constraints are Linear inequalities, written in a unified style Construction of the coupling constraint matrix: Let... To extract the boundary matrix, then Take the adjacent side boundary sequence Given a quantity, Expanding into a two-sided linear inequality, we get
[0133] Both lines are on the right. , It is an all-1 vector with no redundant constant terms; It is a fixed coefficient matrix that does not change with the coordination iteration, only Updated online with the swap sequence. All are constructed recursively from offline identification coefficients, requiring no online fitting. The rolling optimization module uses the effective set method or interior point method to solve the problem. ; Conflict constraint handling: Hard constraints, such as lower temperature limit ≥ 290℃ and actuator physical limit; Soft constraints (rate upper and lower limits, coupling constraints) introduce slack variables. Relaxed to "original constraint" ",Add penalty to the objective function , If the value is a large positive number, and there is no solution, the constraints are relaxed in order of priority: first the coupling constraint, then the rate constraint, until a feasible solution is found; if there is still no solution, a safety fallback is triggered: the flow rate is taken as the upper limit of the anti-condensation cycle, and the heat tracing is taken as the rated power. After one cycle, the value is re-evaluated to ensure that there is a feasible solution in any cycle and the temperature is not lower than 290℃. (4) First element execution and coordination-driven: Only the first element is taken
[0134] The system is decomposed into three sets of parameters: the flow rate setpoint is sent to the flow regulating valve and the frequency converter of the circulating pump; the heat tracing power setpoint is sent to the thyristor power regulator; and the valve opening setpoint is sent to the electric actuator. These three types of actuators act on the thermal process of the heat collection pipes in this area, driving the actual temperature tracking. The remaining sequences are discarded, and the data is collected again in the next cycle and calculated according to step (2). Shift the time domain forward and then solve; (5) Inter-regional coordination, iteration, and convergence guarantee: Before distribution, communication is conducted with adjacent regions via a communication bus. Exchange the boundary temperature prediction sequences, and substitute the adjacent sequences as time-varying boundary conditions. and add coupling constraints .set up For the first The adjacent boundary sequence used in this round of solution is initially set to the latest adjacent exchange sequence. For the adjacent side The newly solved and returned boundary temperature prediction sequence is iterated in the order of solution-exchange-resolve, with three convergence mechanisms: (a) Damping exchange: , A value of 0.5 is preferred to form a compression mapping; Depend on Online calculation; (b) Termination condition: The maximum value of the difference between the two boundary boundaries is less than (Preferred temperature: 0.5℃) or the number of iterations reaches (Preferably 3-5 times) then stop; (c) Custody execution: to achieve If convergence is still not achieved, the latest solution element is executed, and residual mismatches are compensated by rolling feedback; The damping factor compresses the boundary mismatch by a fixed ratio of less than 1 in each round. The mismatch sequence is monotonically decreasing and has a lower bound of zero, so it will inevitably enter the threshold within a finite number of iterations. After the iteration terminates, the first element is executed to make the boundary temperature of adjacent regions consistent and avoid thermal stress.
[0135] Step 5: Configure multi-level temperature hard protection logic independently outside the model predictive control loop as a safety redundancy mechanism.
[0136] Step 6: During the online operation phase, integrate the above modules into a complete closed-loop control system.
[0137] Example 6 This invention is based on a generative adversarial network-driven method for molten salt anti-condensation and safe start-up / shutdown control, specifically implemented according to the following steps: Step 1: Construct a full-system multi-region thermodynamic simulation model of the coupling between the main plant pipeline, the mirror field branch collector pipes, the mirror field main collector header, the transmission pipeline section between the main plant and the mirror field, and the valves; Step 1 is implemented in the following steps: To address the differences in thermal characteristics across different regions, heat balance equations were established for the main plant pipeline, the mirror field branch collector pipes, the mirror field main collector header, the transmission pipeline section between the main plant and the mirror field, and the valves. The mirror field branch adopts a fine-grained multi-node model, which includes an axial segmentation module, a radial layering module, a node energy balance module, and an adjacent node thermal coupling module. The axial segmentation module divides the heat collection tube into multiple calculation units along its length. The radial layering module sets up a molten salt main layer, a molten salt near-wall layer, an inner wall layer, a middle wall layer, and an outer wall layer within each axial calculation unit. The node energy balance module calculates the heat absorption, heat release, heat conduction, and convective heat transfer processes of each node layer. The adjacent node thermal coupling module is used to calculate the molten salt transport heat transfer and axial heat conduction of the tube wall between adjacent axial units, describing the molten salt convective heat transfer inside the tube, the axial heat conduction of the tube wall, and the solar radiation absorption process. The main path of the mirror field adopts a large-diameter thick-walled tube model. The large-diameter thick-walled tube model includes a thick-walled radial heat conduction module, a tube wall heat capacity module, a temperature-related physical property parameter module, and a temperature difference constraint module. The thick-walled radial heat conduction module divides the tube wall from the inside to the outside into an inner wall layer, a middle wall layer, and an outer wall layer, and calculates the heat transfer between each layer according to the cylindrical wall heat conduction equation. The tube wall heat capacity module is used to characterize the heat storage hysteresis of the thick-walled tube during the heating and cooling processes. The temperature-related physical property parameter module is used to correct the thermal conductivity, specific heat capacity, and density of the tube wall material at different temperatures. The temperature difference constraint module is used to calculate the temperature difference between the inner and outer walls and the maximum radial temperature difference of the tube wall, describing the temperature change response under large thermal inertia. For the main plant piping area, a low heat loss model is established considering the low heat dissipation environment indoors. The low heat loss model includes a solar absorption module, a selective absorption coating radiation module, a vacuum interlayer heat transfer module, a glass outer tube convection and radiation heat dissipation module, and a support component thermal bridge loss module. The solar absorption module calculates the effective solar radiation heat received by the collector tube; the selective absorption coating radiation module calculates the radiative heat transfer between the outer surface of the absorber tube and the glass outer tube; the vacuum interlayer heat transfer module describes the radiative heat transfer within the vacuum annulus; the glass outer tube convection and radiation heat dissipation module calculates the natural or forced convection heat dissipation and long-wave radiation heat dissipation between the glass outer tube and the environment; and the support component thermal bridge loss module calculates the additional heat loss caused by local structures such as supports and end connectors. A long-distance heat dissipation model is established for the pipeline section, incorporating ambient wind speed and thermal resistance of the insulation layer. The long-distance heat dissipation model includes a pipeline axial discretization module, a friction loss module, and an environmental boundary condition module. The pipeline axial discretization module divides the long-distance pipeline into multiple continuous pipe sections. The friction loss module calculates the heat released by each pipe section to the environment through the insulation layer. The environmental boundary condition module introduces ambient temperature, wind speed, and surface convective heat transfer coefficient to correct for heat dissipation intensity under different external conditions. A local heat transfer enhancement model for the narrowing and throttling flow in the valve area is established. This model includes an inlet contraction section module, a throat throttling section module, an outlet expansion section module, and an anti-condensation constraint module. The inlet contraction section module calculates the velocity change and local acceleration effect of molten salt entering the narrowing region. The throat throttling section module calculates the flow velocity, Reynolds number, and local heat transfer coefficient at the minimum cross-section of the narrowing. The outlet expansion section module calculates the flow recovery and local backflow effects after expansion. The anti-condensation constraint module determines whether the minimum temperature of molten salt in the narrowing region is lower than the solidification risk threshold and feeds this result back to the safe start-up and shutdown control process. Through joint simulation, key thermodynamic state quantities, including molten salt temperature, pipe wall temperature distribution and maximum temperature difference, inner wall temperature, outer wall temperature, and molten salt heating rate, are calculated in real time for each region.
[0138] Step 2: Construct and train an AC-GANs start trajectory generator with multiple regions and multiple constraints offline, making full use of the conditional generation mechanism of the auxiliary classifier; In step 2, the auxiliary classifier uses five types of area labels—mirror field branch, mirror field main line, main plant, conveying pipeline section, and valve—as input conditions, forcing the generator to learn and distinguish the thermal characteristics of different areas. For the generation of the initial temperature rise trajectory, the mirror field branch uses the inlet molten salt temperature, the initial temperature field of the pipe wall, and the predicted solar irradiance as inputs, with the thermal stress safety threshold and the upper limit of the outlet temperature of 542℃ as hard constraints. Through the discrimination of the auxiliary classifier, the generator is forced to learn to use a step-by-step slow preheating in the low-temperature section to uniformly heat the pipe wall temperature, and actively slow down the heating rate when approaching the upper limit of operation, generating the optimal temperature rise curve for rapid response to irradiation. The conveying pipeline section is constrained by the minimum temperature along the line not being lower than the safety lower limit, generating a conveying trajectory that takes into account both anti-condensation and energy saving. The valve area is constrained by the maximum allowable temperature difference at the diameter reduction point, generating a gradually changing temperature trajectory to avoid thermal shock. During the offline training phase, a startup sample library was first constructed using a fine-grained multi-node model, a large-diameter thick-walled pipe model, a low heat loss model, a long-distance heat dissipation model, and a reduced-diameter throttling local heat transfer enhancement model. The sample library covers the startup process under different ambient temperatures, solar irradiance, cold salt tank outlet temperature, initial pipe wall temperature, molten salt flow rate, pipe length, pipe diameter, insulation heat loss, and reduced-diameter throttling structure parameters. First, the generator receives the startup condition vector:
[0139]
[0140] in, To start the initial molten salt temperature, The initial temperature of the inner wall of the pipe. The initial temperature of the outer wall of the pipe. For ambient temperature, Solar irradiance, For initial traffic, For the length of the pipeline, For pipe diameter, For heat loss per unit length, For a safety lower limit, a temperature range of 285–295°C can be used, with 290°C being preferred. The maximum allowable temperature difference of the pipe wall, This is a random noise vector used to enable the generator to generate multiple sets of candidate trajectories under the same operating conditions. For parameters Generator network, The candidate trajectories output by the generator; The candidate trajectories output by the generator are represented as follows:
[0141] Set the trajectory for the candidate temperatures output by the generator; The candidate mass flow trajectory output by the generator, in units of; The candidate auxiliary heating power trajectory output by the generator; The candidate heating rate trajectory output by the generator; For time; After the generator outputs candidate trajectories, they enter the physical constraint layer, which modifies the candidate curves to obtain physically feasible trajectories.
[0142] The trajectory is corrected by the physical constraint layer; This is a physical constraint correction function; The candidate trajectories output by the generator; This is the current startup condition vector; The corrected trajectory is represented as follows:
[0143] To set the trajectory for the corrected temperature, The corrected mass flow trajectory. The corrected auxiliary heating power trajectory. This is the corrected heating rate trajectory; The discriminator distinguishes the corrected trajectory:
[0144] The safe startup probability or confidence score output by the discriminator has a value range of 0 to 1. For parameters The discriminator; The trajectory is corrected by the physical constraint layer; This is the current startup condition vector; The auxiliary classifier outputs the following operating condition categories:
[0145] This refers to the start-up operating condition category; For parameters Auxiliary classifier; The trajectory is corrected by the physical constraint layer; This is the current startup condition vector; During training, the generator parameters are adjusted in reverse based on the discriminant loss, classification loss, physical constraint loss, and smoothing loss. Total loss function:
[0146] Total loss function of the generator; : Adversarial loss, used to make the generated trajectory approximate the actual safe start trajectory distribution; : Auxiliary classification loss, used to ensure that the generated trajectory conforms to the corresponding startup condition category; Physical constraint loss is used to penalize unsafe conditions such as temperatures below the minimum, excessive heating rates, excessive pipe wall temperature differences, and excessive temperature drops along the pipe. Trajectory smoothing loss is used to reduce abrupt changes in temperature, flow rate, and heating power. : Classification loss weight coefficient; Physical constraint loss weighting coefficient; : Smoothing loss weighting coefficient; Physical constraint loss:
[0147] Physical constraint loss; Minimum temperature constraint penalty coefficient; : Pipe wall temperature difference constraint penalty coefficient; : Maximum heating rate constraint penalty coefficient; Minimum heating rate constraint penalty coefficient; The lowest permissible temperature for the entire system is 290℃. Temperature after physical constraint layer correction; Temperature difference between the inside and outside of the pipe wall; : Maximum allowable temperature difference of the pipe wall; : Heating rate after physical constraint layer correction; : Maximum permissible heating rate; Minimum allowable heating rate; Smoothing loss:
[0148] Trajectory smoothing loss; The corrected temperature at the current moment; Temperature corrected from the previous control cycle; Temperature after correction from the previous two control cycles; Adversarial loss function:
[0149] Adversarial loss on the generator side; The probability that the discriminator considers the corrected generated trajectory to be true. : Calculate the average loss value of all samples in the current batch; The auxiliary classification loss is in the form of the cross-entropy loss function:
[0150] : Auxiliary classification loss; Auxiliary classifier for trajectory determination Belongs to the real category The probability of; Enter the specified working condition category label, either in real or conditional form. : Input the trajectory of the auxiliary classifier.
[0151] In step 2, The startup temperature rise trajectory is a set of candidate trajectories output by the generator. The generator generates candidate startup temperature rise trajectories based on the initial molten salt temperature, pipe wall temperature, ambient temperature, solar irradiance, initial flow rate, pipe diameter, and pipe length. The physical constraint layer is based on the lower limit of molten salt anti-condensation safety, the maximum heating rate, and the maximum temperature difference of the pipe wall. Make corrections to obtain When the local molten salt temperature is between 290℃ and 310℃, the heating rate is 0.3℃ / min to 0.8℃ / min; when the local molten salt temperature is between 310℃ and 400℃, the heating rate can be increased to 0.8℃ / min to 1.5℃ / min; when the local molten salt temperature exceeds 400℃, the heating rate is controlled at 1.0℃ / min to 2.0℃ / min; the discriminator... Conduct a secure boot trustworthiness evaluation, and select those that meet security constraints and have high trustworthiness. As the output of the starting temperature rise trajectory; The optimal heating curve is the best trajectory selected from multiple candidate start-up trajectories output by the generator under the same operating conditions. Below, multiple sets of random noise vectors are input. This causes the generator to output multiple sets of candidate trajectories. Each candidate trajectory is corrected by a physical constraint layer to obtain the corresponding... The discriminator outputs the safe start probability or confidence score for each set of trajectories: A comprehensive evaluation was conducted, taking into account start-up time, temperature safety margin, pipe wall temperature difference, and the smoothness of the heating rate. Items with high safety and reliability, temperatures not lower than 290℃, and pipe wall temperature differences not exceeding [a certain threshold] were selected. The heating rate does not exceed the limit, and the changes in flow rate and heating power are relatively smooth. , as the optimal temperature rise curve; The delivery trajectory corresponds to the mass flow trajectory in the generator output. and the temperature trajectory related to pipe length, pipe diameter, and heat loss per unit length. After being corrected by the physical constraint layer, it forms and ; In the training sample library, the long-distance heat dissipation model will include the pipe length. Pipe diameter Heat loss per unit length Initial flow Ambient temperature Factors such as these are incorporated into the sample. After the generator is trained, it no longer calculates point-by-point according to a certain conveying temperature drop formula. Instead, it generates temperature settings, flow rates, and auxiliary heating trajectories that match the conveying process based on these input parameters. The auxiliary classifier outputs the operating condition category.
[0152] The auxiliary classifier uses five types of area labels: mirror field branch, mirror field main path, main plant, conveying pipeline section, and valve. When the trajectory corresponds to the conveying pipeline section area, This indicates the temperature setting trajectory of the delivery pipeline section. This represents the mass flow rate trajectory during the conveying process. This indicates the trajectory of auxiliary heating power required during the conveying process; The temperature trajectory after physical constraint layer correction corresponds to the gradually varying temperature trajectory. Furthermore, it limits its mutations through a smoothing loss, which is used during training:
[0153] The generator first outputs candidate temperature setpoint trajectories. The physical constraint layer corrects it to Ensure the temperature is not lower than 290℃, and that the heating rate and pipe wall temperature difference do not exceed limits. During training, [the following measures are taken]. By penalizing abrupt changes in the temperature trajectory, the generator gradually outputs a smoother temperature trajectory, ultimately yielding... This refers to a gradually changing temperature trajectory.
[0154] Step 3: Build and train the AC-GANs halt trajectory generator offline. Similarly, use the region category as an auxiliary classification condition to generate differentiated heat preservation strategies for different regions.
[0155] In step 3, The AC-GANs shutdown trajectory generator takes into account the predicted ambient temperature and shutdown duration, with a hard constraint of a minimum safe molten salt temperature of 250℃. For different areas: the mirror field branch and main mirror field generate matching cooling curves based on their respective heat dissipation characteristics and thermal inertia, allowing natural cooling within a safe range; the transport pipe section has high heat loss and freezing risk, so an intermittent heat tracing intervention trajectory is generated, allowing the molten salt inside the pipe to cool naturally within a safe range, with short-term heat replenishment intervention only initiated when the lower limit is reached; the valve area, as a weak point at the diameter reduction, generates a priority insulation trajectory, initiating heat tracing maintenance at the initial stage of shutdown to prevent preferential solidification at the diameter reduction point leading to blockage; the main plant pipeline area utilizes the relatively stable ambient temperature inside the plant to generate a passive insulation strategy, significantly reducing electric heat tracing energy consumption. Each area's trajectory is strictly penalized by a discriminator for results that cause the molten salt temperature inside the pipe to fall below the safe lower limit, while rewarding the generation result with the lowest total energy consumption, thereby minimizing plant power consumption during shutdown while ensuring no freezing. like Figure 1 As shown. The shutdown trajectory generator uses a conditional depthwise convolution generation structure, with the data flow from bottom to top, and consists of four components: (1) Input module: The input consists of three parts, namely the operating condition vector, the region category label, and the random noise vector. The operating condition vector represents the numerical vector of the operating conditions at the time of shutdown. The components include the initial temperature of molten salt, the initial temperature of the inner wall of the pipe, the initial temperature of the outer wall of the pipe, the ambient temperature, the initial mass flow rate, the pipe length, the pipe diameter, the heat loss per unit length, and the lower limit temperature of safety. Maximum allowable temperature difference of pipe wall The input is used as a deterministic condition for trajectory generation; the region category label is the region category of the pipe segment to which the trajectory to be generated belongs, taking one of five categories: mirror field branch, mirror field main road, main plant, conveying pipe segment, and valve; the label is encoded individually or embedded and mapped into a continuous label vector before participating in the fusion, serving as a condition constraint to distinguish the characteristics of trajectories in different regions; random noise vector. : Random vectors sampled according to a standard normal distribution. Under the same working condition vector and region category label, different noise samples correspond to different candidate trajectories, providing multiple sets of candidate samples for subsequent screening; (2) Input fusion node: After the working condition vector, label vector and random noise vector are aligned in dimension, they are fused into a unified initial feature tensor by adding element by element and then fed into the first convolutional basic unit. This node completes the injection of condition information and random information. (3) Convolutional feature basic unit stack: The fused features pass through n concatenated and repeated convolutional feature basic units in sequence. Each basic unit contains three layers in sequence: 3×3 convolutional layer (3×3Conv): performs local convolution operation on time series features to extract local change features such as trend and curvature of trajectory; Normalization layer (Norm): Normalizes the convolution output, suppresses feature value drift, and stabilizes gradient propagation; ReLU activation layer: Outputs features with piecewise linear and nonlinear mapping, providing nonlinear expressive power; in The value is a positive integer, set according to the trajectory sampling length and feature dimension; multiple layers of basic units are stacked sequentially to form the feature extraction backbone of the generator; (4) Upsampling layer: Located after the stack of basic convolutional units, it upsamples the feature sequence along the time dimension, expands the low temporal resolution feature sequence to the set sampling length of the stopping trajectory, and completes the temporal reconstruction from coarse-scale features to full-length trajectory.
[0156] In step 3, (1) The matched cooling curve is used during the shutdown phase, and the generator outputs candidate temperature trajectories. But at this time The cooling rate trajectory is obtained after correction by the physical constraint layer. and ; Input the initial molten salt temperature at shutdown, initial inner wall temperature of the pipe, initial outer wall temperature of the pipe, ambient temperature, current flow rate, pipe length, pipe diameter, heat loss per unit length, safety lower limit, and maximum allowable temperature difference of the pipe wall into the operating condition vector. The generator outputs candidate stopping trajectories. ,in Indicates the candidate cooling curve. Indicates the candidate cooling rate, the physical constraint layer for Make corrections to ensure the corrected version The temperature should not be lower than 290℃, and the cooling rate, pipe wall temperature difference, and friction temperature drop should meet safety requirements. The discriminator should determine whether the cooling trajectory closely approximates the actual safe shutdown trajectory and meets the safe shutdown requirements. Output as a matching cooling curve; (2) The auxiliary heating power trajectory in the generator output corresponding to the intermittent heat tracing intervention trajectory After correction by the physical constraint layer, the following is obtained. ; Safety lower limit Initial molten salt temperature Ambient temperature Heat loss per unit length Pipeline length Pipe diameter Initial flow Maximum allowable temperature difference between pipe wall and pipe wall Input generator. Generator output: candidate auxiliary heating power trajectory. In the scenario of shutdown insulation, This indicates the change in heat tracing power over time. The physical constraint layer, based on the lower limit of temperature safety, the upper limit of heat tracing power, the rate of temperature change, and the smoothness requirements, [determines / imposes / adjusts] the [temperature] power. Revised to When the generator predicts the temperature trajectory Or the corrected temperature trajectory When the temperature approaches 290℃, the physical constraint layer increases the corresponding time period. When the temperature has a sufficient safety margin, reduce or shut off. This ultimately forms an intermittent heat tracing intervention trajectory of opening-maintaining-closing-reopening; (3) Priority insulation trajectory is used to determine the insulation priority of different areas at different times. The auxiliary classifier maps the trajectory to five types of area labels: mirror field branch, mirror field main road, main plant, conveying pipeline section, and valve. For each type of area, the generator will generate a corresponding control trajectory based on its temperature change, heat loss, flow rate and auxiliary heating requirements. If a certain area Closer or the area Larger, pipeline length Longer, flow Lower, required If the value is higher, the area is considered a higher priority for insulation.
[0157] The generator outputs candidate trajectories under various operating conditions. The physical constraint layer was corrected to obtain Auxiliary classifier Identify the region category corresponding to the trajectory, and classify the regions accordingly. , , , , , and close The degree of heat preservation determines the priority order of regional insulation, with higher priority areas receiving more auxiliary heating power. Insulation time and flow rate maintenance strategy; (4) Passive insulation strategy is used to determine whether active heat tracing can be not started immediately after shutdown, and the system waste heat, insulation layer and low flow circulation can be used to maintain a safe temperature. Initial molten salt temperature Initial temperature of the inner wall of the pipe Initial temperature of the outer wall of the pipe Ambient temperature Pipeline length Pipe diameter Heat loss per unit length Initial flow and safety lower limit Input generator, generator outputs candidate temperature trajectories Candidate mass flow trajectory and candidate auxiliary heating power trajectory If the candidate trajectory output by the generator can still meet the requirements of a temperature not lower than 290℃, a pipe wall temperature difference not exceeding the limit, and a cooling rate not too fast when no or minimal auxiliary heating power is used, then the trajectory is determined to be a passive heat preservation candidate strategy, and the physical constraint layer modifies it to... The discriminator outputs a security and reliability score. .like If the value is greater than 0.75, then a passive insulation strategy will be output.
[0158] Step 4: Construct a multi-region coordinated model predictive controller (MPC) as a lower-level actuator to receive the target temperature trajectory output in real time by the AC-GANs generator. In step 4, Each of the five areas—mirror field branch, mirror field main road, main plant, conveying pipeline section, and valve—is equipped with a Model Predictive Controller (MPC), for a total of five MPCs: mirror field branch model predictive controller, mirror field main road model predictive controller, main plant model predictive controller, conveying pipeline section model predictive controller, and valve model predictive controller. Each of these five controllers is independently responsible for temperature tracking and control within its respective area. The five model predictive controllers have the same structure, each consisting of five modules: (1) Input interface module: cycle The initial state is determined by collecting the measured values of the molten salt temperature and the inner / outer wall temperature of the pipe in this area at the beginning. Simultaneously, it receives the corrected trajectory output from the preceding "trajectory generation and physical constraint correction" step. Extract the prediction time domain Temperature components within As a future target trajectory; (2) Prediction model module: Stores the discrete state-space prediction model obtained by linearizing and discretizing the thermal model of this region through the working point, and outputs the predicted temperature sequence based on the initial state and candidate control sequence; (3) Rolling optimization module: Establishes and solves the constrained quadratic programming problem and outputs the optimal control sequence. Its first element is the core solution unit of the controller; (4) Control output module: The first element is decomposed into three set values: molten salt flow rate, electric heating power, and valve opening, which are sent to the flow regulating valve, the circulating pump frequency converter, the electric heater, and the electric actuator, respectively. (5) Coordination and communication module: Exchanges boundary temperature prediction sequences with adjacent area controllers and provides time-varying boundary conditions and inter-regional temperature difference coupling constraints to the rolling optimization module; In step 4, the model prediction controller operates as follows: (1) Status feedback and reference reception: period At the start time, the input interface module collects the measured values of the molten salt temperature and the inner / outer wall temperature of the pipe in this area. This constitutes the initial state. Simultaneously, it receives the corrected trajectory output from the preceding trajectory generation and physical constraint correction steps. Extract the prediction time domain The internal temperature component is used as the future target trajectory, denoted as... ; (2) Model prediction, time-domain preservation and feedback correction: The optimization variable is taken in the control time domain. Control sequence within , These are the molten salt flow rate, electric heating power, and valve opening, respectively, maintaining the following rule: when Time to take That is, control quantity is maintained Unchanged; accordingly The full-time control sequence is ,in To maintain the extended matrix: block rows ( Only in block columns Place Unit block The remaining blocks are all zero blocks, that is... The former Each block row forms a unit matrix, and each subsequent block row repeats the last block row, so that... Recursion for initial values:
[0159] (disturbance) Constant within the time domain, its cumulative terms are incorporated. ), to obtain the predicted temperature sequence , , Depend on Recursive construction, ensuring that the recursion is performed within the specified timeframe. Then proceed according to the rules; Feedback correction: Calculate prediction error (The difference between the measured value and the predicted value for the current moment in the previous period), correct the prediction sequence. , , Optimal choice 1; matrix form , Substitute the corrected sequence into the objective function and temperature constraints of step (3), and save the result in this cycle. For calculation in the next cycle This correction anchors the prediction to the measured value for each cycle, eliminating steady-state error. (3) Constrained rolling optimization and calculation of optimal control sequence: The objective function is
[0160] The tracked item is the corrected predicted value. The reference value is the future target trajectory received in step (1). ,Right now The two cannot be used interchangeably, and there are four categories of constraints: lower limit of temperature. Upper and lower limits of heating rate Actuator limit , Inter-regional boundary coupling constraints (The adjacent boundary sequence is provided by step (5)); Calculation process: Set up the difference (displacement) matrix The main diagonal of the block is The second diagonal of the block is , ,but , This represents the actual control quantity executed in the previous cycle; [record / denote] ( indivual), ( (one), definition Substituting and expanding, we get the standard quadratic form. ,in
[0161]
[0162] All other constraints are Linear inequalities, written in a unified style Construction of the coupling constraint matrix: Let... To extract the boundary matrix, then Take the adjacent side boundary sequence Given a quantity, Expanding into a two-sided linear inequality, we get
[0163] Both lines are on the right. , It is an all-1 vector with no redundant constant terms; It is a fixed coefficient matrix that does not change with the coordination iteration, only Updated online with the swap sequence. All are constructed recursively from offline identification coefficients, requiring no online fitting. The rolling optimization module uses the effective set method or interior point method to solve the problem. ; Conflict constraint handling: Hard constraints, such as lower temperature limit ≥ 290℃ and actuator physical limit; Soft constraints (rate upper and lower limits, coupling constraints) introduce slack variables. Relaxed to "original constraint" ",Add penalty to the objective function , If the value is a large positive number, and there is no solution, the constraints are relaxed in order of priority: first the coupling constraint, then the rate constraint, until a feasible solution is found; if there is still no solution, a safety fallback is triggered: the flow rate is taken as the upper limit of the anti-condensation cycle, and the heat tracing is taken as the rated power. After one cycle, the value is re-evaluated to ensure that there is a feasible solution in any cycle and the temperature is not lower than 290℃. (4) First element execution and coordination-driven: Only the first element is taken
[0164] The system is decomposed into three sets of parameters: the flow rate setpoint is sent to the flow regulating valve and the frequency converter of the circulating pump; the heat tracing power setpoint is sent to the thyristor power regulator; and the valve opening setpoint is sent to the electric actuator. These three types of actuators act on the thermal process of the heat collection pipes in this area, driving the actual temperature tracking. The remaining sequences are discarded, and the data is collected again in the next cycle and calculated according to step (2). Shift the time domain forward and then solve; (5) Inter-regional coordination, iteration, and convergence guarantee: Before distribution, communication is conducted with adjacent regions via a communication bus. Exchange the boundary temperature prediction sequences, and substitute the adjacent sequences as time-varying boundary conditions. and add coupling constraints .set up For the first The adjacent boundary sequence used in this round of solution is initially set to the latest adjacent exchange sequence. For the adjacent side The newly solved and returned boundary temperature prediction sequence is iterated in the order of solution-exchange-resolve, with three convergence mechanisms: (a) Damping exchange: , A value of 0.5 is preferred to form a compression mapping; Depend on Online calculation; (b) Termination condition: The maximum value of the difference between the two boundary boundaries is less than (Preferred temperature: 0.5℃) or the number of iterations reaches (Preferably 3-5 times) then stop; (c) Custody execution: to achieve If convergence is still not achieved, the latest solution element is executed, and residual mismatches are compensated by rolling feedback; The damping factor compresses the boundary mismatch by a fixed ratio of less than 1 in each round. The mismatch sequence is monotonically decreasing and has a lower bound of zero, so it will inevitably enter the threshold within a finite number of iterations. After the iteration terminates, the first element is executed to make the boundary temperature of adjacent regions consistent and avoid thermal stress.
[0165] Step 5: Configure multi-level temperature hard protection logic independently outside the model predictive control loop as a safety redundancy mechanism.
[0166] Step 5 is implemented in the following steps: When the measured outlet temperature of the collector tube reaches the first-level warning threshold of 542℃, the collector power is forcibly reduced. The first-level warning reduces the collector power to 60%~80% of the rated collector power, and the reduction is 20%~40%. When the second-level alarm threshold of 545℃ is reached, the reduction is further increased. The second-level alarm reduces the collector power to 30%~50% of the rated collector power, that is, a cumulative reduction of 50%~70%. When the emergency state threshold of 547℃ is reached, heating is directly cut off and emergency cooling is initiated. When the temperature drops, a hysteresis range is set to prevent control oscillation. This hard protection logic ensures that even if the optimized trajectory generated by artificial intelligence deviates under extreme operating conditions, the actual operating state of the system will not exceed the equipment safety red line.
[0167] Step 6: During the online operation phase, integrate the above modules into a complete closed-loop control system.
[0168] Step 6 is implemented in the following steps: The system first collects real-time data on molten salt temperature, pipe wall temperature, ambient temperature, solar irradiance, ambient wind speed, molten salt mass flow rate, and valve opening in each region using sensors. Then, the collected data is input into the AC-GANs trajectory generator, which generates target temperature trajectories for each region in real time based on the current operating conditions and region category. The target trajectories are then passed to the Model Predictive Controller (MPC) to calculate and execute the optimal control sequence—the third step in the MPC's operation. Simultaneously, multi-level temperature hard protection logic continuously monitors critical temperatures. This multi-level temperature hard protection logic includes the following protection... Protection Strategy: When the measured outlet temperature of the collector tube reaches the first-level warning threshold of 542℃, the collector power is reduced to 70% of the rated collector power; when it reaches the second-level alarm threshold of 545℃, the collector power is reduced to 40% of the rated collector power; when it reaches the emergency state threshold of 547℃, heating is directly cut off and emergency cooling is initiated, switching to a high-flow-rate cold salt circulation to rapidly cool the collector tube; after each level of protection action, the collector power is allowed to be gradually restored only when the measured temperature falls back to a hysteresis range of 3℃ to 5℃ below the trigger threshold, to prevent control oscillation; protection actions are triggered under abnormal operating conditions. The entire control process forms a four-layer closed-loop control architecture of data acquisition → trajectory generation → MPC control execution → safety protection.
Claims
1. A method for molten salt anti-condensation and safe start-up / shutdown control based on generative adversarial networks, characterized in that, The specific steps are as follows: Step 1: Build a full-system multi-area heat collector pipe-molten salt coupled thermodynamic simulation model, including the main plant pipeline, mirror field branch heat collector pipe, mirror field main heat collector header, main plant-mirror field transmission pipeline section and valves; Step 2: Based on the auxiliary classifier condition generation mechanism, construct and train the multi-region, multi-constraint AC-GANs start trajectory generator offline; Step 3: Build and train the AC-GANs halt trajectory generator offline, and output differentiated heat preservation strategies; Step 4: Build a multi-regional coordinated model predictive controller; Step 5: Configure multi-level temperature hard protection logic; Step 6: Achieve complete closed-loop control.
2. The molten salt anti-condensation and safe start-stop control method based on generative adversarial networks according to claim 1, characterized in that, Step 1 is implemented in the following steps: To address the differences in thermal characteristics across different regions, heat balance equations were established for the main plant pipeline, the mirror field branch collector pipes, the mirror field main collector header, the transmission pipeline section between the main plant and the mirror field, and the valves. The mirror field branch adopts a fine-grained multi-node model, which includes an axial segmentation module, a radial layering module, a node energy balance module, and an adjacent node thermal coupling module. The axial segmentation module divides the heat collection tube into multiple calculation units along its length. The radial layering module sets up a molten salt main layer, a molten salt near-wall layer, an inner wall layer, a middle wall layer, and an outer wall layer within each axial calculation unit. The node energy balance module calculates the heat absorption, heat release, heat conduction, and convective heat transfer processes of each node layer. The adjacent node thermal coupling module is used to calculate the molten salt transport heat transfer and axial heat conduction of the tube wall between adjacent axial units, describing the molten salt convective heat transfer inside the tube, the axial heat conduction of the tube wall, and the solar radiation absorption process. The main path of the mirror field adopts a large-diameter thick-walled tube model. The large-diameter thick-walled tube model includes a thick-walled radial heat conduction module, a tube wall heat capacity module, a temperature-related physical property parameter module, and a temperature difference constraint module. The thick-walled radial heat conduction module divides the tube wall from the inside to the outside into an inner wall layer, a middle wall layer, and an outer wall layer, and calculates the heat transfer between each layer according to the cylindrical wall heat conduction equation. The tube wall heat capacity module is used to characterize the heat storage hysteresis of the thick-walled tube during the heating and cooling processes. The temperature-related physical property parameter module is used to correct the thermal conductivity, specific heat capacity, and density of the tube wall material at different temperatures. The temperature difference constraint module is used to calculate the temperature difference between the inner and outer walls and the maximum radial temperature difference of the tube wall, describing the temperature change response under large thermal inertia. For the main plant piping area, a low heat loss model is established considering the low heat dissipation environment indoors. The low heat loss model includes a solar absorption module, a selective absorption coating radiation module, a vacuum interlayer heat transfer module, a glass outer tube convection and radiation heat dissipation module, and a support component thermal bridge loss module. The solar absorption module calculates the effective solar radiation heat received by the collector tube; the selective absorption coating radiation module calculates the radiative heat transfer between the outer surface of the absorber tube and the glass outer tube; the vacuum interlayer heat transfer module describes the radiative heat transfer within the vacuum annulus; the glass outer tube convection and radiation heat dissipation module calculates the natural or forced convection heat dissipation and long-wave radiation heat dissipation between the glass outer tube and the environment; and the support component thermal bridge loss module calculates the additional heat loss caused by local structures such as supports and end connectors. A long-distance heat dissipation model is established for the pipeline section, incorporating ambient wind speed and thermal resistance of the insulation layer. The long-distance heat dissipation model includes a pipeline axial discretization module, a friction loss module, and an environmental boundary condition module. The pipeline axial discretization module divides the long-distance pipeline into multiple continuous pipe sections. The friction loss module calculates the heat released by each pipe section to the environment through the insulation layer. The environmental boundary condition module introduces ambient temperature, wind speed, and surface convective heat transfer coefficient to correct for heat dissipation intensity under different external conditions. A local heat transfer enhancement model for the narrowing and throttling flow in the valve area is established. This model includes an inlet contraction section module, a throat throttling section module, an outlet expansion section module, and an anti-condensation constraint module. The inlet contraction section module calculates the velocity change and local acceleration effect of molten salt entering the narrowing region. The throat throttling section module calculates the flow velocity, Reynolds number, and local heat transfer coefficient at the minimum cross-section of the narrowing. The outlet expansion section module calculates the flow recovery and local backflow effects after expansion. The anti-condensation constraint module determines whether the minimum temperature of molten salt in the narrowing region is lower than the solidification risk threshold and feeds this result back to the safe start-up and shutdown control process. Through joint simulation, key thermodynamic state quantities, including molten salt temperature, pipe wall temperature distribution and maximum temperature difference, inner wall temperature, outer wall temperature, and molten salt heating rate, are calculated in real time for each region.
3. The molten salt anti-condensation and safe start-stop control method based on generative adversarial networks according to claim 1, characterized in that, In step 2, the auxiliary classifier uses five types of area labels—mirror field branch, mirror field main road, main plant, conveying pipeline section, and valve—as input conditions, forcing the generator to learn and distinguish the thermal characteristics of different areas. For the generation of the starting temperature rise trajectory, the mirror field branch uses the inlet molten salt temperature, the initial temperature field of the pipe wall, and the predicted solar irradiance as inputs, and uses the thermal stress safety threshold and the upper limit of the outlet temperature of 542℃ as hard constraints. Through the discrimination of the auxiliary classifier, the generator is forced to learn to use a step-by-step slow preheating in the low-temperature section to uniformly heat the pipe wall temperature, and actively slow down the heating rate when approaching the upper limit of operation, generating the optimal temperature rise curve for rapid response to irradiation. The conveying pipeline section is constrained by the minimum temperature along the line not being lower than the safety lower limit, generating a conveying trajectory that takes into account both anti-condensation and energy saving. The valve area is constrained by the maximum allowable temperature difference at the diameter reduction point, generating a gradually changing temperature trajectory to avoid thermal shock. During the offline training phase, a startup sample library was first constructed using a fine-grained multi-node model, a large-diameter thick-walled pipe model, a low heat loss model, a long-distance heat dissipation model, and a reduced-diameter throttling local heat transfer enhancement model. The sample library covers the startup process under different ambient temperatures, solar irradiance, cold salt tank outlet temperature, initial pipe wall temperature, molten salt flow rate, pipe length, pipe diameter, insulation heat loss, and reduced-diameter throttling structure parameters. First, the generator receives the startup condition vector: in, To start the initial molten salt temperature, The initial temperature of the inner wall of the pipe. The initial temperature of the outer wall of the pipe. For ambient temperature, Solar irradiance, For initial traffic, For the length of the pipeline, For pipe diameter, For heat loss per unit length, To ensure a safe lower limit, The maximum allowable temperature difference of the pipe wall, This is a random noise vector used to enable the generator to generate multiple sets of candidate trajectories under the same operating conditions. For parameters Generator network, The candidate trajectories output by the generator; The candidate trajectories output by the generator are represented as follows: Set the trajectory for the candidate temperatures output by the generator; The candidate mass flow trajectory output by the generator, in units of; The candidate auxiliary heating power trajectory output by the generator; The candidate heating rate trajectory output by the generator; For time; After the generator outputs candidate trajectories, they enter the physical constraint layer, which modifies the candidate curves to obtain physically feasible trajectories. The trajectory is corrected by the physical constraint layer; This is a physical constraint correction function; The candidate trajectories output by the generator; This is the current startup condition vector; The corrected trajectory is represented as follows: To set the trajectory for the corrected temperature, The corrected mass flow trajectory. The corrected auxiliary heating power trajectory. This is the corrected heating rate trajectory; The discriminator distinguishes the corrected trajectory: The safe startup probability or confidence score output by the discriminator has a value range of 0 to 1. For parameters The discriminator; The trajectory is corrected by the physical constraint layer; This is the current startup condition vector; The auxiliary classifier outputs the following operating condition categories: This refers to the start-up operating condition category; For parameters Auxiliary classifier; The trajectory is corrected by the physical constraint layer; This is the current startup condition vector; During training, the generator parameters are adjusted in reverse based on the discriminant loss, classification loss, physical constraint loss, and smoothing loss. Total loss function: Total loss function of the generator; : Adversarial loss, used to make the generated trajectory approximate the actual safe start trajectory distribution; : Auxiliary classification loss, used to ensure that the generated trajectory conforms to the corresponding startup condition category; Physical constraint loss is used to penalize unsafe conditions such as temperatures below the minimum, excessive heating rates, excessive pipe wall temperature differences, and excessive temperature drops along the pipe. Trajectory smoothing loss is used to reduce abrupt changes in temperature, flow rate, and heating power. : Classification loss weight coefficient; Physical constraint loss weighting coefficient; : Smoothing loss weighting coefficient; Physical constraint loss: Physical constraint loss; Minimum temperature constraint penalty coefficient; : Pipe wall temperature difference constraint penalty coefficient; : Maximum heating rate constraint penalty coefficient; Minimum heating rate constraint penalty coefficient; The lowest permissible temperature for the entire system is 290℃. Temperature after physical constraint layer correction; Temperature difference between the inside and outside of the pipe wall; : Maximum allowable temperature difference of the pipe wall; : Heating rate after physical constraint layer correction; : Maximum permissible heating rate; Minimum allowable heating rate; Smoothing loss: Trajectory smoothing loss; The corrected temperature at the current moment; Temperature corrected from the previous control cycle; Temperature after correction from the previous two control cycles; Adversarial loss function: Adversarial loss on the generator side; The probability that the discriminator considers the corrected generated trajectory to be true; : Calculate the average loss value of all samples in the current batch; The auxiliary classification loss is in the form of the cross-entropy loss function: : Auxiliary classification loss; Auxiliary classifier for trajectory determination Belongs to the real category The probability of; Enter the specified working condition category label, either in real or conditional form. : Input the trajectory of the auxiliary classifier.
4. The molten salt anti-condensation and safe start-stop control method based on generative adversarial networks according to claim 3, characterized in that, In step 2, The startup temperature rise trajectory is a set of candidate trajectories output by the generator. The generator generates candidate startup temperature rise trajectories based on the initial molten salt temperature, pipe wall temperature, ambient temperature, solar irradiance, initial flow rate, pipe diameter, and pipe length. ; The physical constraint layer is based on the lower limit of molten salt anti-condensation safety, the maximum heating rate, and the maximum temperature difference of the pipe wall. Make corrections to obtain When the local molten salt temperature is between 290℃ and 310℃, the heating rate is 0.3℃ / min to 0.8℃ / min; when the local molten salt temperature is between 310℃ and 400℃, the heating rate can be increased to 0.8℃ / min to 1.5℃ / min; when the local molten salt temperature exceeds 400℃, the heating rate is controlled at 1.0℃ / min to 2.0℃ / min; the discriminator... Conduct a secure boot trustworthiness evaluation, and select those that meet security constraints and have high trustworthiness. As the output of the starting temperature rise trajectory; The optimal heating curve is the best trajectory selected from multiple candidate start-up trajectories output by the generator under the same operating conditions. Below, multiple sets of random noise vectors are input. This causes the generator to output multiple sets of candidate trajectories. Each candidate trajectory is corrected by a physical constraint layer to obtain the corresponding... The discriminator outputs the safe start probability or confidence score for each set of trajectories: A comprehensive evaluation was conducted, taking into account start-up time, temperature safety margin, pipe wall temperature difference, and the smoothness of the heating rate. Items with high safety and reliability, temperatures not lower than 290℃, and pipe wall temperature differences not exceeding [a certain threshold] were selected. The heating rate does not exceed the limit, and the changes in flow rate and heating power are relatively smooth. , as the optimal temperature rise curve; The delivery trajectory corresponds to the mass flow trajectory in the generator output. and the temperature trajectory related to pipe length, pipe diameter, and heat loss per unit length. After being corrected by the physical constraint layer, it forms and ; In the training sample library, the long-distance heat dissipation model will include the pipe length. Pipe diameter Heat loss per unit length Initial flow Ambient temperature Factors such as these are incorporated into the sample. After the generator is trained, it no longer calculates point-by-point according to a certain conveying temperature drop formula. Instead, it generates temperature settings, flow rates, and auxiliary heating trajectories that match the conveying process based on these input parameters. The auxiliary classifier outputs the operating condition category. The auxiliary classifier uses five types of area labels: mirror field branch, mirror field main path, main plant, conveying pipeline section, and valve. When the trajectory corresponds to the conveying pipeline section area, This indicates the temperature setting trajectory of the delivery pipeline section. This represents the mass flow rate trajectory during the conveying process. This indicates the trajectory of auxiliary heating power required during the conveying process; The temperature trajectory after physical constraint layer correction corresponds to the gradually varying temperature trajectory. Furthermore, it limits its mutations through a smoothing loss, which is used during training: The generator first outputs candidate temperature setpoint trajectories. The physical constraint layer corrects it to Ensure the temperature is not lower than 290℃, and that the heating rate and pipe wall temperature difference do not exceed limits. During training, [the following measures are taken]. By penalizing abrupt changes in the temperature trajectory, the generator gradually outputs a smoother temperature trajectory, ultimately yielding... This refers to a gradually changing temperature trajectory.
5. The molten salt anti-condensation and safe start-stop control method based on generative adversarial networks according to claim 4, characterized in that, In step 3 The AC-GANs shutdown trajectory generator takes into account the predicted ambient temperature and shutdown duration. A hard constraint is a minimum safe molten salt temperature of 250℃. For different areas: the mirror field branch and main mirror field generate matching cooling curves based on their respective heat dissipation characteristics and thermal inertia, allowing natural cooling within a safe range; the transport pipe section has high heat loss and freezing risk, so an intermittent heat tracing intervention trajectory is generated, allowing the molten salt inside the pipe to cool naturally within a safe range, with short-term heat replenishment intervention only initiated when the lower limit is reached; the valve area, as a weak point at the diameter reduction, generates a priority insulation trajectory, initiating heat tracing maintenance at the initial stage of shutdown to prevent preferential solidification at the diameter reduction point leading to blockage; the main plant pipeline area utilizes the relatively stable ambient temperature inside the plant to generate a passive insulation strategy, significantly reducing electric heat tracing energy consumption. Each area's trajectory is strictly penalized by a discriminator for results that cause the molten salt temperature inside the pipe to fall below the safe lower limit, while rewarding the generation result with the lowest total energy consumption, thereby minimizing plant power consumption during shutdown while ensuring no freezing. The shutdown trajectory generator uses a conditional deep convolutional generation structure, with the data flow from bottom to top, and consists of four components: (1) Input module: The input consists of three parts, namely the operating condition vector, the region category label, and the random noise vector. The operating condition vector represents the numerical vector of the operating condition at the time of shutdown. The components include the initial temperature of molten salt, the initial temperature of the inner wall of the pipe, the initial temperature of the outer wall of the pipe, the ambient temperature, the initial mass flow rate, the pipe length, the pipe diameter, the heat loss per unit length, and the lower safety limit temperature. Maximum allowable temperature difference of pipe wall The input is used as a deterministic condition for trajectory generation; the region category label is the region category of the pipe segment to which the trajectory to be generated belongs, taking one of five categories: mirror field branch, mirror field main road, main plant, conveying pipe segment, and valve; the label is encoded individually or embedded and mapped into a continuous label vector before participating in the fusion, serving as a condition constraint to distinguish the characteristics of trajectories in different regions; random noise vector. : Random vectors sampled according to a standard normal distribution. Under the same working condition vector and region category label, different noise samples correspond to different candidate trajectories, providing multiple sets of candidate samples for subsequent screening; (2) Input fusion node: After the working condition vector, label vector and random noise vector are aligned in dimension, they are fused into a unified initial feature tensor by adding element by element and then fed into the first convolutional basic unit. This node completes the injection of condition information and random information. (3) Convolutional feature basic unit stack: The fused features pass through n concatenated and repeated convolutional feature basic units in sequence. Each basic unit contains three layers in sequence: 3×3 convolutional layer: performs local convolution operation on time series features to extract local change features such as trend and curvature of trajectory; Normalization layer: Normalizes the convolution output, suppresses feature value drift, and stabilizes gradient propagation; ReLU activation layer: Outputs features with piecewise linear and nonlinear mapping, providing nonlinear expressive power; in The value is a positive integer, set according to the trajectory sampling length and feature dimension; multiple layers of basic units are stacked sequentially to form the feature extraction backbone of the generator; (4) Upsampling layer: Located after the stack of basic convolutional units, it upsamples the feature sequence along the time dimension, expands the low temporal resolution feature sequence to the set sampling length of the stopping trajectory, and completes the temporal reconstruction from coarse-scale features to full-length trajectory.
6. The molten salt anti-condensation and safe start-stop control method based on generative adversarial networks according to claim 5, characterized in that, In step 3 (1) The matched cooling curve is used during the shutdown phase, and the generator outputs candidate temperature trajectories. But at this time The cooling rate trajectory is obtained after correction by the physical constraint layer. and ; Input the initial molten salt temperature at shutdown, initial inner wall temperature of the pipe, initial outer wall temperature of the pipe, ambient temperature, current flow rate, pipe length, pipe diameter, heat loss per unit length, safety lower limit, and maximum allowable temperature difference of the pipe wall into the operating condition vector. The generator outputs candidate stopping trajectories. ,in Indicates the candidate cooling curve. Represents the candidate cooling rate, and the physical constraint layer... Make corrections to ensure the corrected version The temperature should not be lower than 290℃, and the cooling rate, pipe wall temperature difference, and friction temperature drop should meet safety requirements. The discriminator should determine whether the cooling trajectory closely approximates the actual safe shutdown trajectory and meets the safe shutdown requirements. Output as a matching cooling curve; (2) The auxiliary heating power trajectory in the generator output corresponding to the intermittent heat tracing intervention trajectory After correction by the physical constraint layer, the following is obtained. ; Safety lower limit Initial molten salt temperature Ambient temperature Heat loss per unit length Pipeline length Pipe diameter Initial flow Maximum allowable temperature difference between pipe wall and pipe wall Input generator, generator output candidate auxiliary heating power trajectory In the scenario of shutdown insulation, This indicates the change in heat tracing power over time. The physical constraint layer, based on the lower limit of temperature safety, the upper limit of heat tracing power, the rate of temperature change, and the smoothness requirements, [determines / imposes / adjusts] the [temperature] power. Revised to When the generator predicts the temperature trajectory Or the corrected temperature trajectory When the temperature approaches 290℃, the physical constraint layer increases the corresponding time period. When the temperature has a sufficient safety margin, reduce or shut off. This ultimately forms an intermittent heat tracing intervention trajectory of opening-maintaining-closing-reopening; (3) Priority insulation trajectory is used to determine the insulation priority of different areas at different times. The auxiliary classifier maps the trajectory to five types of area labels: mirror field branch, mirror field main road, main plant, conveying pipe section, and valve. For each type of area, the generator will generate the corresponding control trajectory based on its temperature change, heat loss, flow rate and auxiliary heating requirements. The generator outputs candidate trajectories under various operating conditions. The physical constraint layer was corrected to obtain Auxiliary classifier Identify the region category corresponding to the trajectory, and classify the regions accordingly. , , , , , and close The degree of heat preservation determines the priority order of regional insulation, with higher priority areas receiving more auxiliary heating power. Insulation time and flow maintenance strategy; (4) Passive insulation strategy is used to determine whether active heat tracing can be not started immediately after shutdown, and the system waste heat, insulation layer and low flow circulation can be used to maintain a safe temperature. Initial molten salt temperature Initial temperature of the inner wall of the pipe Initial temperature of the outer wall of the pipe Ambient temperature Pipeline length Pipe diameter Heat loss per unit length Initial flow and safety lower limit Input generator, generator outputs candidate temperature trajectories Candidate mass flow trajectory and candidate auxiliary heating power trajectory If the candidate trajectory output by the generator can still meet the requirements of a temperature not lower than 290℃, a pipe wall temperature difference not exceeding the limit, and a cooling rate not too fast when no or minimal auxiliary heating power is used, then the trajectory is determined to be a passive heat preservation candidate strategy, and the physical constraint layer modifies it to... The discriminator outputs a security and reliability score. ,like If the value is greater than 0.75, then a passive insulation strategy will be output.
7. The molten salt anti-condensation and safe start-stop control method based on generative adversarial networks according to claim 1, characterized in that, In step 4 Each of the five areas—mirror field branch, mirror field main road, main plant, conveying pipeline section, and valve—is equipped with a model predictive controller, resulting in a total of five model predictive controllers: mirror field branch model predictive controller, mirror field main road model predictive controller, main plant model predictive controller, conveying pipeline section model predictive controller, and valve model predictive controller. Each of these five controllers is independently responsible for temperature tracking and control within its respective area. The five model predictive controllers have the same structure, each consisting of five modules: (1) Input interface module: cycle The initial state is determined by collecting the measured values of the molten salt temperature and the inner / outer wall temperature of the pipe in this area at the beginning. Simultaneously, it receives the corrected trajectory output from the preceding "trajectory generation and physical constraint correction" step. Extract the prediction time domain Temperature components within As a future target trajectory; (2) Prediction model module: Stores the discrete state-space prediction model obtained by linearizing and discretizing the thermal model of this region through the working point, and outputs the predicted temperature sequence based on the initial state and candidate control sequence; (3) Rolling optimization module: Establishes and solves the constrained quadratic programming problem and outputs the optimal control sequence. Its first element is the core solution unit of the controller; (4) Control output module: The first element is decomposed into three set values: molten salt flow rate, electric heating power, and valve opening, which are sent to the flow regulating valve, the circulating pump frequency converter, the electric heater, and the electric actuator, respectively. (5) Coordination and communication module: exchange boundary temperature prediction sequences with adjacent area controllers, and provide time-varying boundary conditions and inter-regional temperature difference coupling constraints to the rolling optimization module.
8. The molten salt anti-condensation and safe start-stop control method based on generative adversarial networks according to claim 7, characterized in that, In step 4, the model prediction controller operates as follows: (1) Status feedback and reference reception: period At the start time, the input interface module collects the measured values of the molten salt temperature and the inner / outer wall temperature of the pipe in this area. This constitutes the initial state. Simultaneously, it receives the corrected trajectory output from the preceding trajectory generation and physical constraint correction steps. Extract the prediction time domain The internal temperature component is used as the future target trajectory, denoted as... ; (2) Model prediction, time-domain preservation and feedback correction: The optimization variable is taken in the control time domain. Control sequence within , These are the molten salt flow rate, electric heating power, and valve opening, respectively, maintaining the following rule: when Time to take That is, control quantity is maintained Unchanged; accordingly The full-time control sequence is ,in To maintain the extended matrix: block rows ( Only in block columns Place Unit block The remaining blocks are all zero blocks, that is... The former Each block row forms a unit matrix, and each subsequent block row repeats the last block row, so that... Recursion for initial values: Disturbance Constant within the time domain, its cumulative terms are incorporated. , to obtain the predicted temperature sequence , , Depend on Recursive construction, ensuring that the recursion is performed within the specified timeframe. Then proceed according to the rules; Feedback correction: Calculate prediction error Correct the predicted sequence , , Optimal choice 1; matrix form , Substitute the corrected sequence into the objective function and temperature constraints of step (3), and save the result in this cycle. For calculation in the next cycle This correction anchors the prediction to the measured value for each cycle, eliminating steady-state error. (3) Constrained rolling optimization and calculation of optimal control sequence: The objective function is The tracked item is the corrected predicted value. The reference value is the future target trajectory received in step (1). ,Right now The two cannot be used interchangeably, and there are four categories of constraints: lower limit of temperature. Upper and lower limits of heating rate Actuator limit , ; Inter-regional boundary coupling constraints ; Calculation process: Let the difference matrix be set. The main diagonal of the block is The second diagonal of the block is , ,but , This represents the actual control quantity executed in the previous cycle; [record / denote] ( indivual), ,definition Substituting and expanding, we get the standard quadratic form. ,in All other constraints are Linear inequalities, written in a unified style Construction of coupling constraint matrix: Let To extract the boundary matrix, then Take the adjacent side boundary sequence Given a quantity, Expanding into a two-sided linear inequality, we get Both lines are on the right. , It is an all-1 vector with no redundant constant terms; It is a fixed coefficient matrix that does not change with the coordination iteration, only Updated online with the swap sequence. All are constructed recursively from offline identification coefficients, requiring no online fitting. The rolling optimization module uses the effective set method or interior point method to solve the problem. ; Conflict constraint handling: hard constraints, lower temperature limit ≥ 290℃, actuator physical limit; soft constraints introduce slack variables. Relaxed to "original constraint" ",Add penalty to the objective function , If the value is a large positive number, and there is no solution, the constraints are relaxed in order of priority: first the coupling constraint, then the rate constraint, until a feasible solution is found; if there is still no solution, a safety fallback is triggered: the flow rate is taken as the upper limit of the anti-condensation cycle, and the heat tracing is taken as the rated power. After one cycle, the value is re-evaluated to ensure that there is a feasible solution in any cycle and the temperature is not lower than 290℃. (4) First element execution and coordination-driven: Only the first element is taken The system is decomposed into three sets of parameters: the flow rate setpoint is sent to the flow regulating valve and the circulating pump frequency converter; the heat tracing power setpoint is sent to the thyristor power regulator; and the valve opening setpoint is sent to the electric actuator. These three types of actuators act on the thermal process of the heat collection pipes in this area, driving the actual temperature tracking. The remaining sequences are discarded, and the data is collected again in the next cycle and calculated according to step (2). Shift the time domain forward and then solve; (5) Inter-regional coordination, iteration, and convergence guarantee: Before distribution, communication is conducted with adjacent regions via a communication bus. Exchange the boundary temperature prediction sequences, and substitute the adjacent sequences as time-varying boundary conditions. and add coupling constraints ,set up For the first The adjacent boundary sequence used in this round of solution is initially set to the latest adjacent exchange sequence. For the adjacent side The newly solved and returned boundary temperature prediction sequence is iterated in the order of solution-exchange-resolve, with three convergence mechanisms: (a) Damping exchange: This constitutes a compression mapping; Depend on Online calculation; (b) Termination condition: The maximum value of the difference between the two boundary boundaries is less than Or the number of iterations reaches That is, terminate; (c) Custody execution: to achieve If convergence is still not achieved, the latest solution element is executed, and residual mismatches are compensated by rolling feedback; The damping factor compresses the boundary mismatch by a fixed ratio of less than 1 in each round. The mismatch sequence is monotonically decreasing and has a lower bound of zero, so it will inevitably enter the threshold within a finite number of rounds. After the iteration terminates, the first element is executed to ensure that the boundary temperature of adjacent regions is consistent, thus avoiding thermal stress.
9. The molten salt anti-condensation and safe start-stop control method based on generative adversarial networks according to claim 1, characterized in that, Step 5 is implemented in the following steps: When the measured outlet temperature of the collector tube reaches the first-level warning threshold of 542℃, the collector power is forcibly reduced. The first-level warning reduces the collector power to 60%~80% of the rated collector power, and the reduction is 20%~40%. When the second-level alarm threshold of 545℃ is reached, the reduction is further increased. The second-level alarm reduces the collector power to 30%~50% of the rated collector power, that is, a cumulative reduction of 50%~70%. When the emergency state threshold of 547℃ is reached, heating is directly cut off and emergency cooling is initiated. When the temperature drops, a hysteresis range is set to prevent control oscillation. This hard protection logic ensures that even if the optimized trajectory generated by artificial intelligence deviates under extreme operating conditions, the actual operating state of the system will not exceed the equipment safety red line.
10. The molten salt anti-condensation and safe start-stop control method based on generative adversarial networks according to claim 1, characterized in that, Step 6 is implemented in the following steps: The system first collects real-time data on molten salt temperature, pipe wall temperature, ambient temperature, solar irradiance, ambient wind speed, molten salt mass flow rate, and valve opening in each area using sensors. Then, the collected data is input into the AC-GANs trajectory generator, which generates target temperature trajectories for each area based on the current operating conditions and area category. The target trajectories are then passed to the model predictive controller, which calculates and executes the optimal control sequence—the third step in the model predictive controller's operation. Simultaneously, multi-level temperature hard protection logic continuously monitors critical temperatures. This multi-level temperature hard protection logic includes the following protection strategy: when the measured collector tube outlet temperature reaches the first-level warning threshold of 542℃, the collector power is reduced to 70% of the rated collector power. When the secondary alarm threshold of 545℃ is reached, the heat collection power is reduced to 40% of the rated heat collection power; when the emergency state threshold of 547℃ is reached, heating is directly cut off and emergency cooling is initiated, switching to a high-flow-rate cold salt circulation to rapidly cool the heat collection tubes; after each level of protection action, the heat collection power is allowed to be restored step by step only when the measured temperature falls back to the hysteresis range of 3℃ to 5℃ below the trigger threshold, to prevent control oscillation; under abnormal operating conditions, the protection action is triggered, and the entire control process forms a four-layer closed-loop control of data acquisition → trajectory generation → MPC control execution → safety protection.
