An energy-saving coupling control method for a direct air-cooled auxiliary molten salt heat storage system

Abstract

The application discloses an energy-saving coupling control method for a direct air-cooling auxiliary molten salt heat storage system, comprising the following steps: S1, constraint setting of working condition boundary description; S2, parameter self-consistency of dynamic image building; S3, state reliable reconstruction of multi-source data integration; S4, hierarchical weight arrangement of target layer construction; S5, forward-looking decision kernel rolling solution; S6, self-adaptive identification gain setting mechanism; S7, coupling resource arrangement and collaborative distribution; S8, execution layer decoupling driving protection logic injection; S9, operation evaluation closed-loop knowledge sedimentation update. Through accurate working condition boundary description and dynamic modeling, multi-source data fusion state estimation and forward-looking optimization decision, the application establishes a self-adaptive coupling control system with constraints, effectively solves the problems of rough modeling, poor adaptability and multi-loop coupling conflict of the traditional control strategy, and realizes efficient energy saving, stable and reliable operation and strategy continuous evolution of the direct air-cooling auxiliary molten salt heat storage system under complex working conditions.

Description

Technical Field

[0001] This invention relates to an energy-saving coupling control method for a direct air-cooled assisted molten salt thermal energy storage system, belonging to the field of energy storage and utilization technology. Background Technology

[0002] With the global energy crisis and environmental pollution becoming increasingly severe, the development of efficient and sustainable energy storage and utilization technologies has become particularly important. In the field of thermal energy storage, molten salt thermal storage systems, due to their high thermal density and long storage cycle, have been widely used in solar power generation, industrial waste heat recovery, and other applications. However, in systems that couple direct air cooling with molten salt thermal storage, traditional control strategies have many shortcomings, which seriously affect energy utilization efficiency and system stability, becoming a key problem that urgently needs to be solved.

[0003] Existing molten salt thermal energy storage systems and their coupled control methods typically rely on experience-based proportional-integral-derivative (PID) control or simple thermal balance control. These methods often exhibit poor robustness and adaptability when facing complex dynamic changes and external disturbances. On the one hand, traditional control methods neglect the impact of complex internal and external disturbances, lacking accurate modeling and real-time optimization of the system's dynamic characteristics, resulting in energy waste and control instability. On the other hand, existing technologies fail to effectively handle nonlinear effects in the heat transfer process under complex operating conditions, cannot flexibly adjust control strategies according to actual operating conditions, and lack fine coordination and regulation of key equipment such as air-cooled fans and molten salt pumps, leading to insufficient energy utilization and the system's inability to achieve optimal performance in changing environments. Furthermore, traditional methods struggle to balance safety and economy when dealing with equipment constraints, multi-loop coupling, and sudden changes in operating conditions, easily leading to risks such as equipment overload and temperature exceeding limits, and cannot iteratively optimize control strategies through continuous accumulation and learning of operating data.

[0004] Therefore, there is an urgent need for an energy-saving coupled control method that can perform fine modeling, dynamic optimization, and adaptive adjustment of direct air-cooled assisted molten salt thermal storage systems, in order to minimize energy consumption, improve operating efficiency, ensure stable and reliable system operation, and achieve continuous evolution of control strategies. Summary of the Invention

[0005] The purpose of this invention is to provide an energy-saving coupling control method for a direct air-cooled assisted molten salt thermal energy storage system, in order to solve the problems of coarse system modeling, poor adaptability of control strategies, multi-loop coupling conflicts, and imprecise energy management in the prior art.

[0006] To solve the above-mentioned technical problems, the present invention adopts the following technical solution: an energy-saving coupling control method for a direct air-cooled assisted molten salt thermal storage system, comprising the following steps: S1. Defining and Establishing Constraints for Operating Condition Boundaries: Layer disturbance factors according to time scale and establish statistical characteristics, establish equipment operating condition boundary constraints that include hard and soft constraints, unify the energy balance caliber, conduct observability analysis of measurement points, and introduce virtual sensing. S2. Dynamic profiling parameter self-consistency: The system is divided into multiple dynamic partitions, a dynamic model containing nonlinearity and inertia is established, the model parameters are corrected online, and re-identification is triggered by monitoring the energy closure residual. S3, Multi-source data embellishment and reliable state reconstruction: Time alignment, delay compensation and anomaly weighting are performed on the measurement data, and the system state is estimated and the uncertainty is output using extended Kalman filtering; S4. Hierarchical weight arrangement of target layer construction: Construct a multi-objective function that includes energy consumption, tracking deviation, action frequency and boundary approximation penalty, perform dimensionless normalization and dynamically adjust weights, and set hard constraints to have absolute priority and soft constraints to be graded down. S5. Forward-looking decision kernel rolling solution: Rolling optimization is performed using a multi-rate model predictive control framework, combined with hot start and fault rollback mechanisms, to filter and freeze the changes of the online updated model parameters. S6. Adaptive identification gain tuning mechanism: Select key parameters through sensitivity analysis, identify them online using recursive least squares method with forgetting factor, map them to controller parameters, and set an abnormal freeze rollback mechanism. S7. Coupling Resource Orchestration and Cooperative Allocation: Introduces multi-loop coupling conflicts into constraint optimization, allocates resources based on the principle of minimum marginal cost, and uses consistency supervision quantities to suppress oscillations; S8. Execution layer decoupling-driven protection logic injection: Establish nonlinear mapping and decoupling protection logic for the execution mechanism, prioritize arbitration with hard constraints, and feed back the execution status to the estimator and optimizer; S9. Operational evaluation closed-loop knowledge accumulation and update: Hierarchical recording of operational data and events, root cause analysis, and version management and rollback of control strategies and parameters after playback verification.

[0007] The aforementioned energy-saving coupling control method for a direct air-cooled assisted molten salt thermal energy storage system, wherein step S1 specifically includes: S1.1. Ambient temperature, wind speed and direction, intraday irradiance variation, unit load fluctuation, and cold end available resource fluctuation are taken as disturbance factors and stratified according to time scales of seconds, minutes and hours, and defined as three categories: measurable, estimable and unobservable. At the same time, statistical characteristics including mean, variance, correlation time and typical transition amplitude are established. S1.2. Take the heat exchange section, pump valve, bypass, fan / fin section, and heating section as key objects, and write the allowable working range of their material temperature limit, phase window, and pressure difference limit as hard constraints. Write the allowable working range of their temperature approaching the boundary and fan speed approaching the limit as soft constraints. Give the dimensions of each constraint, the consistency of the sampling period, and the out-of-bounds handling strategy including penalty, saturation, and degradation. S1.3. Clarify the conventions for positive and negative heat values, flow rate benchmarks, enthalpy calculation benchmarks, rules for including pressure drop losses, and the placement of auxiliary power consumption in the energy balance, and solidify these into a data dictionary; S1.4 Map the layout of measurement points for temperature, differential pressure, flow rate, valve position, speed, and electrical power to a set of state variables, perform rank / sensitivity analysis, and if unobservable states or weakly controllable quantities are found, introduce alternative variables, virtual sensors, or supplementary measurement points at the strategy layer.

[0008] The aforementioned energy-saving coupling control method for a direct air-cooled assisted molten salt thermal energy storage system, wherein step S2 specifically includes: S2.1. Divide the system into several dynamic partitions. For each partition, select the partition average temperature, equivalent heat capacity, equivalent heat transfer coefficient, and flow hysteresis state as state variables. The selection principle is that the state energy can be closed and energy is conserved, it can be deduced from the measurement, and it is sensitive to the manipulated quantity. At the same time, the time constant and update cycle of each state are clearly defined. S2.2 The heat transfer coefficient changes with flow velocity, the physical properties change with temperature, the air-side heat transfer is coupled with the fan speed and the ambient wind speed, and the transport delay introduced by the pipeline / heat exchanger is treated as nonlinearity and inertia, and represented by differentiable functions. The pure delay is implemented using first-order / multi-order hysteresis or discrete shift registers. S2.3 First, use historical data to offline fit a set of initial values ​​for heat transfer / resistance / heat capacity under multiple representative working conditions. Then, during operation, use a recursive method to slowly correct the drift rate and add a regularization term, and set a confidence threshold: when sensor drift or anomaly occurs, pause the adaptive process and fall back to the conservative parameter set. S2.4 Calculate the energy closure residual for each sampling period, and allocate the residual to three channels according to the confidence level: measurement deviation, unmodeled thermal loss, and physical property error. Set up residual statistical monitoring. If the residual is biased for a long time, trigger the re-identification or measurement point calibration process.

[0009] The aforementioned energy-saving coupling control method for a direct air-cooled assisted molten salt thermal energy storage system, wherein step S3 specifically includes: S3.1. Perform timestamp correction and resampling to a uniform step size for each measurement point, and establish a channel delay model that includes sensor response, communication buffer, and filter group delay; perform delay compensation for key closed-loop quantities or use predicted values ​​to enter the controller. S3.2. Identify jump points, deadlocks, drifts, and saturation as data ailments. Perform combined cleaning through physical boundary checks, differential mutation detection, consistency checks, and short-window statistical tests. Assign a confidence weight to each data point so that the estimator can automatically ignore suspicious observations. S3.3. State estimation is performed using an extended Kalman filter under noise and model error conditions. Its discrete-time form is as follows:

[0010] In the formula, Let be the system state vector. Its estimated value, To manipulate the input vector, To measure the output vector, For discrete state transition functions, For measurement function, Let Jacobian be the state transition with respect to the state. Let P be the Jacobian of the measurement function with respect to the state, Q be the process noise covariance, R be the measurement noise covariance, and Kk be the Kalman gain. S3.4 Output State Estimation The uncertainty P or its diagonal variance, including key derivatives such as heat flow, equivalent drag, and energy inventory, as well as the weights of each measurement point, are used by the controller for risk constraints or adaptive parameter tuning.

[0011] The aforementioned energy-saving coupling control method for a direct air-cooled assisted molten salt thermal storage system, wherein step S4 specifically includes: S4.1. Write the power consumption terms of the fan and pump as the convex penalty of the input, write the temperature / heat flow tracking as the deviation penalty, write the frequent actions as the input rate of change penalty, and write the approach to the boundary as the soft constraint relaxation penalty. S4.2. Reduce the priority of wind turbine-related penalty weights when wind speed is high; increase the weight of motion smoothing terms when measurement point reliability is low; increase constraint relaxation penalty when equipment approaches the boundary. S4.3 Normalize each penalty to a dimensionless quantity using a reference value, which includes allowable deviation, rated power, and maximum rate of change, and record the source of each reference value; S4.4. Set hard constraints to always take precedence; soft constraints are degraded in the order of releasing comfort zone type first, then efficiency type, and finally motion smoothness type, and define trigger conditions for each degrade level, including duration, out-of-bounds magnitude, and estimated uncertainty threshold.

[0012] The aforementioned energy-saving coupling control method for a direct air-cooled assisted molten salt thermal storage system, wherein step S5 specifically includes: S5.1 Make the prediction time domain cover the main thermal inertia from minutes to hours, the control time domain shorter than the prediction time domain, and set up a multi-rate mechanism: fast loop handles the details of the actuator, and slow loop handles energy distribution and boundary management; S5.2. Load the indicators and constraints defined in step S4 into the constraint model predictive control framework, using a quadratic form:

[0013] In the formula, Np is the number of prediction steps, Nc is the number of control steps, and ΔU is the control increment sequence. rk+i is the reference trajectory. Let be the weighted L2 norm, Qy, Ru, and S be the weight matrices for output bias, input increment, and soft constraint relaxation, respectively, A, B, and C be the discrete linear model matrices, and dk+i be the equivalent perturbation / modeling error compensation term. For the set of constraint functions that need to be softened, These are slack variables; S5.3. Employ a mechanism that includes hot start, constraint projection, and reverting to the previous feasible control or switching conservative rules in case of failure, and bind the solver iteration count, termination condition, and sampling period. S5.4. Apply amplitude limiting and low-pass filtering to the parameter changes obtained from online identification; set a freeze zone for changes in the key matrix, and do not update if the change is less than the threshold; and perform a feasibility check after each update to ensure that the constraints are met and the gain is normal.

[0014] The aforementioned energy-saving coupling control method for a direct air-cooled assisted molten salt thermal storage system, wherein step S6 specifically includes: S6.1 First, perform sensitivity analysis, select the few parameters that have the greatest impact on key outputs to enter online updates, and reparameterize highly correlated parameters, merging two product terms into one equivalent parameter; S6.2. The recursive least squares method with a forgetting factor is used to correct slowly changing equivalent parameters online. Its standard form is:

[0015] In the formula, For the estimation of the parameter vector to be identified, In order to return to the goal, For the regression vector, The covariance is estimated for the parameters, where λ∈(0,1] is the forgetting factor. For RLS gain; S6.3, will The parameters are mapped to weights, model matrices, and disturbance compensation coefficients available within the controller for parameter tuning, and the physical rationality of the parameters, model stability and controllability are checked simultaneously, as well as the constraint boundaries are recalculated. S6.4 Set the freeze condition to sensor weights that are too low, energy residual statistics that are abnormal, or the solver that is frequently infeasible; hold after triggering. Revert to the most recent health parameter snapshot without changing the settings, and record the event.

[0016] The aforementioned energy-saving coupling control method for a direct air-cooled assisted molten salt thermal energy storage system, wherein step S7 specifically includes: S7.1. Write the coupling conflicts of multiple input pairs with multiple states and power consumption into the objectives and constraints in the form of cross terms, shared constraints, and coupled dynamics. S7.2. Power consumption, near-boundary risk, and motion wear are converted into a unified marginal cost curve that is updated with the operating conditions, and the controller allocates power based on the principle of minimizing marginal cost. S7.3. Represent the collaborative assignment as a constrained optimization model, using the augmented Lagrangian form:

[0017] Its Lagrange function is:

[0018] In the formula, u is the collaborative allocation decision variable. Let h(u) = 0 be the cost term of type j, h(u) = 0 be the equality constraint, g(u) ≤ 0 be the inequality constraint, c(u) be the coupling conflict quantity or residual term to be suppressed, λ and μ be the Lagrange multipliers, and ρ > 0 be the quadratic penalty coefficient. S7.4 Introduce a consistency supervision quantity. When oscillation signs are detected, increase the smoothing weight, shorten the control time domain, or increase damping, and limit the simultaneous large-scale movement of multiple channels.

[0019] The aforementioned energy-saving coupling control method for a direct air-cooled assisted molten salt thermal storage system, wherein step S8 specifically includes: S8.1 Establish valve opening mapping, pump / fan speed / power mapping, and a limiting and dead zone processing model that includes minimum resolution, hysteresis, saturation, and start / stop constraints, and impose slope limits on commands. S8.2 Separate the master control quantity and slave quantity through feedforward cancellation, static decoupling matrix or cascade structure. When a conflict occurs, perform priority arbitration, satisfy hard constraint related loops first, and then process performance loops. S8.3 Make the trigger threshold, hold time, release condition, and regression slope of the protection logic consistent with the sampling period; when protection is triggered, simultaneously tighten the constraints of the upper-level optimizer or switch to the degradation target; S8.4 Real-time monitoring of command-feedback deviation, actuator jamming signs, saturation ratio and cumulative action cost, and writing back valid input feedback to the estimator and optimizer to replace command input for status updates.

[0020] The aforementioned energy-saving coupling control method for a direct air-cooled assisted molten salt thermal energy storage system, wherein step S9 specifically includes: S9.1 Record key states, inputs, constraint occupancy rates, soft constraint relaxation amounts, estimation uncertainties, and solver states on time scales of seconds, minutes, and hours, and label abnormal, protection trigger, parameter freeze, and sensor calibration events. S9.2, Combining energy closure residuals and EKF innovative sequences Statistics and MPC infeasibility frequency are used to perform root cause decomposition of measurement bias, unmodeled loss, parameter drift, disturbance prediction error, and actuator failure, and provide processing actions such as calibration, re-identification, adjustment of noise covariance, or tightening of constraints. S9.3. Perform replay verification on typical scenarios including high disturbance, low load, boundary approach, and sensor anomaly to check whether infeasibility occurs, whether too many protections are triggered, or whether input oscillation occurs. Only after passing the verification should the new weight rules, new parameter initial values, and new thresholds be written to the version library. S9.4 records the version number, change item, applicable working conditions, and rollback path for each change, and incorporates key hyperparameters Q, R, λ, weight matrix, and soft constraint penalties into configuration management, allowing the runtime to roll back to the previous stable version with one click.

[0021] Compared with the prior art, the present invention has at least the following beneficial effects: This invention establishes constraints by characterizing operating condition boundaries, constructs parameter self-consistency by building dynamic profiles, and embellishes state reliability reconstruction by multi-source data. It layers complex disturbances on time scales and establishes hard and soft constraints to build an accurate dynamic model containing nonlinearity and inertia. Combined with data reliability weighting and extended Kalman filtering, it achieves robust perception of system state and its uncertainty, solves the problems of coarse modeling and poor disturbance resistance of traditional methods, and provides a reliable basis for optimized control.

[0022] This invention constructs a multi-objective function that includes energy consumption, tracking, smoothing, and boundary risks by constructing hierarchical weights and using a look-ahead decision kernel for rolling solution. It then employs multi-rate model predictive control for rolling optimization and combines hot start and fault fallback mechanisms to enable the system to dynamically balance multiple objectives, anticipate and address thermal inertia. This avoids the low efficiency and energy waste caused by lag adjustment in traditional PID control under complex constraints, significantly improving energy scheduling and operational efficiency.

[0023] This invention utilizes an adaptive identification gain tuning mechanism and a coordinated allocation of coupled resources. It selects key parameters using sensitivity analysis and corrects them online using a recursive least squares method with a forgetting factor. Simultaneously, it transforms multi-loop coupling conflicts into constraint optimization with a unified marginal cost, enabling the system to automatically adapt to equipment aging and operating condition drift. It also coordinates the allocation of hot and cold resources with the minimum marginal cost, effectively eliminating loop oscillations and improving the system's adaptability and coupled operational stability under varying operating conditions.

[0024] This invention establishes an actuator mapping with nonlinearity and slope constraints and hard constraint priority arbitration by decoupling the execution layer to drive protection logic injection and operation evaluation closed-loop knowledge accumulation and update. Based on root cause analysis and playback verification, the control strategy is versioned and rolled back with one click. This not only ensures the safe and reliable operation of physical equipment, but also realizes the closed-loop iteration and knowledge reuse of the control strategy, which greatly reduces the difficulty of operation and maintenance, failure risk and manual debugging cost.

[0025] The synergistic effect of the above steps in this invention enables efficient global management of direct air cooling and molten salt thermal storage systems under complex operating conditions. Through precise power regulation and equipment protection, it significantly reduces ineffective energy consumption and thermal fatigue damage of equipment such as fans and molten salt pumps, reduces system operating energy consumption, maintenance and downtime costs, extends the service life of core equipment, improves energy utilization efficiency from the source, and achieves significant energy saving, carbon reduction and economic benefits. Attached Figure Description

[0026] Figure 1 This is a schematic diagram of the steps of the present invention.

[0027] The present invention will be further described below with reference to the accompanying drawings and specific embodiments. Detailed Implementation

[0028] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be clearly and completely described below with reference to the accompanying drawings and specific embodiments. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0029] like Figure 1 As shown, an energy-saving coupled control method for a direct air-cooled assisted molten salt thermal energy storage system includes the following steps: S1, Delineation and Constraint of Working Condition Boundaries; S2, Dynamic portrait construction parameters are self-consistent; S3, Reliable reconstruction of multi-source data embellished state; S4. Target layer construction hierarchical weight arrangement; S5, forward-looking decision kernel rolling solution; S6, Adaptive identification gain tuning mechanism; S7, Coupling resource orchestration and collaborative allocation; S8, Injection of execution layer decoupling driver protection logic; S9. Operational evaluation closed-loop knowledge accumulation and update.

[0030] The specific implementation methods for each step are described in detail below.

[0031] Step S1: Characterize and establish constraints for the working condition boundary. This step aims to establish clear boundary conditions and disturbance interfaces for subsequent control and optimization, specifically including: S1.1 Disturbance Characterization and Layering: Identify the main disturbance sources in the system, including ambient temperature, wind speed and direction, intraday irradiance variations, unit load fluctuations, and fluctuations in available cold-end resources. These disturbances are categorized by time scale into second-level, minute-level, and hour-level scales, and the observability of each type of disturbance is defined: measurable (directly measurable by sensors), estimable (indirectly estimable through models), and unobservable (can only be processed according to statistical characteristics). For each type of disturbance, its statistical characteristics are established, including mean, variance, correlation time, and typical transition amplitude, providing a unified disturbance interface for prediction and robust control.

[0032] S1.2 Constraint Establishment: For key components such as heat exchange sections, pumps and valves, bypasses, fan / fin segments, and heating sections, establish constraint expressions for the allowable operating range based on their physical limitations and safety requirements. Constraints are divided into hard constraints and soft constraints. Hard constraints are inviolable boundaries, such as material temperature limits, phase windows, and pressure differential limits. Soft constraints are boundaries that can be touched briefly but at a cost, such as temperatures approaching material boundaries or fan speeds approaching limits. For each constraint, clarify its dimensions and sampling period consistency, and define out-of-bounds handling strategies, including penalties (adding costs to the objective function), saturation (mandatory limit instructions), and degradation (switching to protection mode).

[0033] S1.3 Unified Energy Balance Standards: Establish conventions for heat sign, flow rate benchmarks, reference temperatures and states for enthalpy calculations, clarify whether pressure drop losses are included in the energy balance, and define the allocation of auxiliary power consumption (pump power, fan power, valve losses) in energy audits. Solidify these conventions into a data dictionary to ensure that model identification, online optimization, and performance evaluation use completely consistent energy audit rules.

[0034] S1.4 Observability Analysis: The existing measurement point layout (temperature, differential pressure, flow rate, valve position, speed, electrical power) is mapped to the system state variable set. Through rank analysis and sensitivity analysis, it is determined whether all key states are observable and all manipulated variables are controllable. If unobservable states (such as local heat capacity, lag temperature) or weakly controllable quantities (such as a valve having an extremely weak effect on a certain temperature) are found, alternative variables, virtual sensors, or suggestions for supplementary measurement points are introduced at the strategy level.

[0035] Step S2: Dynamic portrait construction parameter self-consistency This step establishes a dynamic model that can be updated online for the system, namely a "dynamic profile," which specifically includes: S2.1 Dynamic Zoning and State Selection: The air-cooled assisted molten salt thermal energy storage system is decomposed into several dynamic zones, each containing a heat exchange, transport, or energy storage component. State variables characterizing energy accumulation and transfer are selected for each zone, typically including zone average temperature, equivalent heat capacity, equivalent heat transfer coefficient, and flow hysteresis state. The selection of state variables must meet the following conditions: they must close the energy conservation equation, be obtainable through measurement or estimation, and have sufficient sensitivity to manipulated variables. Simultaneously, the time constant and update period of each state variable are clearly defined to avoid numerical ill-conditioning caused by mixing fast and slow processes in the same difference equation.

[0036] S2.2 Nonlinear and Inertial Modeling: Known nonlinear effects and inertial elements in the system are explicitly incorporated into the model structure. For example, the relationship between the heat transfer coefficient and flow velocity is described using piecewise smoothing or exponential functions; the physical properties of molten salt and air as a function of temperature are represented by polynomials or power functions; the wind-side heat transfer coefficient is coupled with the fan speed and ambient wind speed; the transport delay introduced by the pipeline and heat exchanger is implemented using first-order / multi-order hysteresis elements or discrete shift registers. This model provides a sufficiently accurate foundation for subsequent linearization, prediction, and stability analysis.

[0037] S2.3 Initial Parameter Acquisition and Online Correction: Before commissioning, a set of initial values ​​for model parameters, including equivalent heat transfer coefficient, drag coefficient, and heat capacity, are fitted offline using historical data from multiple representative operating conditions. During real-time operation, parameters are slowly corrected using a recursive method. The parameter drift rate is limited and a regularization term is introduced during the correction process to prevent abrupt model changes caused by abnormal data. A confidence threshold is also set: when sensor drift or data anomalies are detected, adaptive updates are paused, and parameters are reverted to a pre-stored conservative parameter set to ensure control safety.

[0038] S2.4 Energy Closure Residual Monitoring: In each sampling period, the energy closure residual is calculated based on the model and measurement data, which is the difference between the input energy and the output energy, as well as the change in stored energy. The residual is assigned to three error channels according to confidence level: measurement bias, unmodeled heat loss, and physical property error. A statistical monitoring mechanism for the residual is established. If the residual shows a long-term bias rather than random fluctuation, the model re-identification or measurement point calibration process is triggered.

[0039] Step S3: Reliable reconstruction of multi-source data embellishment state This step aims to provide the controller with clean, synchronized, and confident state information, specifically including: S3.1 Time Alignment and Delay Compensation: Timestamp correction and resampling are performed on the data from each sensor to unify them to the same step size. A channel delay model is established to address group delays introduced by sensor response time, communication buffering, and filtering. For key variables involved in closed-loop control, delay compensation is implemented, or model-based predicted values ​​are directly used in the controller to avoid instability caused by phase lag.

[0040] S3.2 Data Outlier Removal and Weighting: Identifying and handling common "data outliers," including jump points (isolated erroneous values), stuck data (data remaining unchanged for a long time), drift (slowly deviating from the true value), and saturation (reaching the measurement limit). A combined strategy is employed: physical boundary checks directly mark out-of-range data; differential mutation detection marks outliers in the rate of change; consistency checks mark suspicious observations based on sudden increases in energy balance residuals; and short-window statistical tests (such as MAD or Huber methods) further remove outliers. During the cleaning process, data is not directly deleted; instead, each data point is assigned a confidence weight of 0 to 1, enabling the subsequent state estimator to automatically ignore suspicious observations.

[0041] S3.3 Extended Kalman Filter State Estimation: In the presence of process noise and measurement noise, a discrete-time extended Kalman filter (EKF) is used for state estimation. State prediction uses the nonlinear dynamic model f(·) established in step S2, and measurement updates utilize sensor readings y_k. The filter formula is as follows: State prediction:

[0042] Covariance prediction:

[0043] Kalman gain calculation:

[0044] Status Update:

[0045] in: : System state vector (zone temperature, equivalent heat capacity state, hysteresis state); Its estimated value; Manipulate input vectors (such as a set of controllable variables like pumps, valves, and fans); : Measurement output vector (set of sensor readings); Discrete state transition function (given by the nonlinear dynamic profile you created); Measurement function (a relationship that maps states to measured quantities); : The Jacobian of state transitions with respect to the state; : The Jacobian of the measurement function with respect to the state; P: State estimation error covariance; Q: Process noise covariance (describes unmodeled disturbances / model uncertainties); R: Measurement noise covariance; Kalman gain (determining whether to trust the model or the measurement).

[0046] Step S4: Target layer construction and weight arrangement This step involves constructing a multi-objective optimization problem and dynamically managing its weights and priorities, specifically including: S4.1 Construction of Multi-Objective Penalty Terms: The control objectives are written into the objective function in the form of penalty terms. The power consumption term (electrical energy consumption of fans and pumps) is represented as a convex function of the input quantity, the temperature / heat flow tracking deviation is represented as the squared error, the frequent actions of the actuators are represented as the squared penalty of the input rate of change, and the risk of approaching the constraint boundary is represented as the penalty of the soft constraint slack variable. This forms clear and non-conflicting mathematical objectives.

[0047] S4.2 Dynamic Weight Adjustment: The weights of each penalty item are adjusted in real time according to the operating conditions. When the ambient wind speed is high, the marginal benefit of heat dissipation on the wind side decreases. At this time, the weight of the fan-related penalty is reduced to allow it to slow down appropriately to save energy. When the reliability of the sensor decreases, the weight of the action smoothing item is increased to avoid the controller from generating high-frequency jitter due to noise. When the equipment state approaches the hard constraint boundary, the constraint relaxation penalty is significantly increased to force the optimization solution to move away from the danger zone.

[0048] S4.3 Dimensionless Normalization: To avoid difficulties in weight adjustment due to differences in the dimensions of various targets, a benchmark value is used to normalize each penalty. The benchmark value can be taken from equipment nameplate parameters (such as rated power), operating experience data (such as allowable temperature deviation), or historical data statistical quantiles (such as maximum rate of change). The source of each benchmark value must be recorded to ensure traceability.

[0049] S4.4 Constraint Priority and Degradation Logic: Hard constraints have the highest priority and must never be violated. Soft constraints are categorized into multiple levels based on their urgency: Level 1 is the comfort zone (e.g., the optimal efficiency zone), Level 2 is efficiency-related (allowing for some efficiency sacrifice), and Level 3 is action smoothing-related (allowing for larger actions). When the system cannot simultaneously satisfy all soft constraints due to extreme disturbances, they are released level by level in this order. Clear triggering conditions are defined for each level of degradation, including duration, boundary exceedance magnitude, and state estimation uncertainty threshold, ensuring the system maintains clear decision-making logic even in critical situations.

[0050] Step S5: Rolling solution of the forward-looking decision kernel This step involves online calculation of the optimal control sequence based on model predictive control, specifically including: S5.1 Multi-rate time domain setup: The prediction time domain covers the main thermal inertia of the system, typically ranging from several minutes to several hours, ensuring that dynamic responses can be anticipated; the control time domain can be relatively short to reduce computational load. A multi-rate architecture is also designed: the fast loop operates with shorter cycles to handle the details of actuator instructions; the slow loop operates with longer cycles to handle energy allocation and boundary management, avoiding the explosion of the optimization problem size.

[0051] S5.2 Constrained MPC Quadratic Form Optimization: The multi-objective penalties and constraints defined in step S4 are integrated into a constrained model predictive control problem of the following form:

[0052] satisfy:

[0053] in: Predict the number of steps; : Control the number of steps; ; Reference trajectory (may come from the scheduling layer / running instructions); Weighted L2 norm; Output bias weight matrix; Input the incremental weight matrix; : Soft constraint relaxation weight matrix; A, B, C: Discrete linear model matrix (which can be obtained by linearizing the nonlinear model at the current point); : Equivalent perturbation / modeling error compensation term (which can be provided by the estimator or perturbation observer); : The set of constraint functions that need to be softened (temperature approaching the boundary, pressure difference approaching the upper limit); Slack variables (allow soft constraints to be slightly violated, but will be penalized by S).

[0054] The optimization problem is solved online in each control cycle, applying only the current control increment Δu_k to the system.

[0055] S5.3 Solution Guarantee Mechanism: To meet real-time requirements, a warm-start strategy is adopted, that is, the optimal solution of the previous cycle is shifted as the initial value for the current solution; if the initial value does not meet the constraints, constraint projection is first performed to make it fall into the feasible region; if the solver cannot converge within the specified time, it reverts to the feasible control of the previous step, or switches to a set of preset conservative rules. The number of iterations and convergence criteria of the solver are strictly bound to the sampling period to ensure hard real-time performance.

[0056] S5.4 Parameter Online Update Protection: The model parameter changes identified in step S6 must undergo amplitude limiting and low-pass filtering before being sent to MPC to avoid abrupt changes. For key matrices in the prediction model, a change freeze zone is set; that is, updates are not performed when the change is less than a preset threshold to prevent numerical fluctuations. After each update, a quick check is performed to determine if the optimization problem is still feasible and whether the control gain is abnormal, ensuring the stability of the rolling decision.

[0057] Step S6: Adaptive Identification Gain Tuning Mechanism This step involves online correction of model parameters and mapping them to controller parameter tuning, specifically including: S6.1 Parameter Selection and Reparameterization: Through global sensitivity analysis, a few physical parameters that have the greatest impact on key outputs (such as temperature and heat flux) are selected for online identification, avoiding overfitting and computational burden caused by high-dimensional identification. For strongly correlated parameters (such as heat transfer coefficient K and area A, which often appear as a product term KA), they are merged into an equivalent parameter for identification, improving the reliability of identification.

[0058] S6.2 Online Recursive Least Squares Identification: For the selected slowly changing equivalent parameters, online estimation is performed using the recursive least squares (RLS) method with a forgetting factor.

[0059] in, The vector of parameters to be identified (such as equivalent heat transfer coefficient, drag coefficient, and heat capacity correction factor). The regression objective is constructed based on energy balance or temperature difference-heat flow relationship. It is a regression vector (composed of measured or estimated states). Let be the parametric covariance matrix, and λ∈(0,1] be the forgetting factor. The smaller λ is, the more attention is paid to new data, but the more sensitive the estimation is. This is the RLS gain, used to adjust the magnitude of parameter updates per step.

[0060] S6.3 Parameter Mapping and Verification: Mapping and verifying the identified parameters The mapping is to quantities that can be directly used within the controller, such as updating the coefficients in the MPC model matrices A, B, and C, adjusting the target weights or disturbance compensation terms, etc. Before updating the application, physical rationality verification (whether the parameter signs, numerical ranges, and monotonicity are correct) must be performed, and the controllability and observability of the new model must be quickly verified. At the same time, the relevant constraint boundaries should be recalculated (e.g., tightening the maximum flow constraint simultaneously when the pressure drop coefficient increases).

[0061] S6.4 Abnormal Freeze and Rollback: Set safety monitoring conditions for parameter updates. When abnormal situations such as excessively low sensor weights, abnormal energy residuals, or frequent infeasibility of MPC solving occur, the system will immediately freeze the process. The update maintains the current identification value and automatically rolls back to the most recent parameter snapshot marked "healthy", while recording the event for offline analysis.

[0062] Step S7: Coupling resource orchestration and collaborative allocation This step addresses the coordination issue of multiple loops on both the air-cooling and thermal storage sides, specifically including: S7.1 Coupling Conflict Modeling: Identify coupling conflicts in the system. For example, increasing the flow rate on one side can improve thermal storage efficiency but worsen the air-cooled back pressure, or increasing the fan speed can reduce the cold-end temperature but significantly increase plant power consumption. Incorporate these coupling relationships into the objective function and constraints in the form of cross terms, shared constraints, and coupled dynamics to avoid optimizing for local optima or actions that lead to oscillations.

[0063] S7.2 Unified Marginal Cost Conversion: Energy consumption costs, near-limit risk costs, and equipment wear and tear costs are uniformly converted into a marginal cost curve measured by economic efficiency, and dynamically updated according to operating conditions (such as electricity price and equipment status). In this way, when resources are scarce or disturbances intensify, the controller can dynamically allocate load among loops according to the principle of minimizing marginal costs, replacing rigid switching with fixed priorities.

[0064] S7.3 Augmented Lagrangian Optimization Model: Transforms the cooperative allocation problem into a constrained optimization problem with a quadratic penalty term.

[0065] Its Lagrange function is:

[0066] In the formula, u is the set of controllable inputs (pump speed, valve position, fan speed, etc.). Let h(u) = 0 represent equality constraints (such as mass and energy conservation), g(u) ≤ 0 represent inequality constraints (upper and lower limits, rate limits), c(u) be the coupling conflict quantity or residual term that needs to be specially suppressed, λ be the Lagrange multiplier for equality constraints, μ be the Lagrange multiplier for inequality constraints (satisfying μ ≥ 0), and ρ > 0 be the penalty coefficient. By adding a coupling penalty term to the objective, the optimization solution can be forced to converge towards the overall optimal direction of the system, avoiding single-loop sacrificial optimization.

[0067] S7.4 Oscillation Suppression Supervision: A consistency supervision quantity is introduced to monitor in real time the coupling residual c(u), the correlation of changes in control variables in adjacent periods, and the bias of prediction errors. When oscillation signs are detected (such as periodic fluctuations in a variable or large synchronous back-and-forth adjustments in multiple channels), the supervision mechanism actively intervenes: temporarily increasing the smoothing weight of the input rate of change, shortening the prediction time domain to limit over-looking ahead, or increasing damping (increasing the penalty for Δu), and limiting multiple channels from making large movements simultaneously, thereby avoiding triggering system resonance or instability.

[0068] Step S8: Injection of execution layer decoupling driver protection logic This step transforms the optimized instructions into executable, safe instructions, specifically including: S8.1 Nonlinear Mapping of Actuators: Establish precise mapping relationships for each actuator, including valve opening-flow characteristics and pump / fan speed-power / flow characteristics, while fully considering their physical limitations, such as minimum resolution, hysteresis, saturation characteristics, and start / stop constraints. Simultaneously, apply slope limits to commands to match the rate of change of commands with mechanical and electrical characteristics, preventing high-frequency jitter and shocks.

[0069] S8.2 Decoupling and Priority Arbitration: Using feedforward compensation, static decoupling matrices, or cascade control structures, strongly coupled loops are decoupled into master and slave variables, simplifying control relationships. When multi-loop control commands conflict (e.g., an air-cooled fan requires full speed while protection logic requires speed reduction), priority arbitration is executed: commands from loops related to hard constraints are fully satisfied first, then commands from performance-optimized loops are processed, ensuring absolute safety.

[0070] S8.3 Synchronization of Protection Logic and Optimizer: The trigger threshold, hold time, release condition, and regression ramp rate of the protection function must be coordinated with the sampling period. Once protection is triggered (such as temperature over-limit protection), in addition to executing the protection action locally, the information must also be synchronized to the upper-level optimizer so that it can tighten relevant constraints or switch to a degraded objective function, to prevent the optimizer from continuing to issue instructions that the execution layer "cannot do", and to prevent deep saturation and integrator accumulation.

[0071] S8.4 Execution Status Feedback: Continuously monitor the actual response of the actuator, calculate the deviation between the command and the feedback, and diagnose phenomena such as jamming (persistent deviation and no response), excessive saturation ratio, and abnormal accumulation of action costs. The actual effective input (not the original command) is sent back as the feedback value u to the EKF state estimator and MPC optimizer, ensuring that the model and optimization are based on the actual execution status, thus avoiding control model mismatch due to inadequate execution.

[0072] Step S9: Operational evaluation and knowledge accumulation / update. This step enables long-term performance monitoring and strategy evolution of the control system, specifically including: S9.1 Hierarchical Data Recording and Labeling: Key state variables, input commands, constraint occupancy rates, soft constraint relaxation amounts, state estimation uncertainties, and solver operating status are recorded at multiple time scales, including second, minute, and hourly levels. Simultaneously, various events occurring during operation are automatically labeled, such as anomalies, protection triggers, parameter freezes, and sensor calibrations, providing structured tags for subsequent analysis.

[0073] S9.2 Root Cause Analysis: Combining energy closure residuals and EKF innovative sequences (i.e., measurement prediction errors) Based on the statistical characteristics of [missing information] and the frequency of MPC solution infeasibility, root cause decomposition is performed on the following typical problems: measurement bias, unmodeled loss, parameter drift, disturbance prediction error, and actuator failure. For each type of root cause, recommended actions are automatically provided, such as sensor calibration, triggering model re-identification, adjusting the noise covariance matrices Q and R, and automatically tightening key constraints.

[0074] S9.3 Replay Verification and Strategy Iteration: Typical historical scenarios such as high disturbance, low load, equipment approaching the boundary, and sensor anomalies are extracted for offline replay verification of newly generated control strategies, model parameters, and threshold configurations. Verification indicators include: whether optimization becomes infeasible, whether protection is over-triggered, and whether control input oscillations occur. Only new strategies, parameters, and thresholds that pass all verifications are allowed to be written into the online version control repository, achieving closed-loop iterative evolution.

[0075] S9.4 Version Management and One-Click Rollback: Each change to the control strategy, model parameters, weight matrix, or hyperparameters (such as Q, R, λ) automatically generates a new version, recording the version number, changes, applicable operating conditions, and rollback path. The runtime can roll back to the previous stable version with one click, ensuring that a known reliable configuration can be restored at any time, greatly reducing the risk of online parameter tuning.

[0076] By organically combining the above steps S1 to S9, this method can perform global, adaptive, and highly reliable energy management of the direct air-cooled assisted molten salt thermal storage system, significantly improving the energy-saving effect and stability of the system under complex and variable operating conditions.

[0077] The working principle of one embodiment of the present invention is as follows: In actual use, the present invention identifies and defines the main influencing factors in the system, including ambient temperature, wind speed, intraday irradiance variation, and unit load fluctuations. These disturbance factors are stratified according to time scale (second-level, minute-level, hour-level), and by defining measurable, estimable, and non-observable categories, it helps the system predict possible changes in operating conditions in advance during operation, ensuring that the system can respond quickly and adopt robust control strategies. This process provides a stable data information interface for subsequent optimization decisions by establishing unified statistical characteristics (such as mean, variance, transition amplitude, etc.).

[0078] The system uses a dynamic profiling model to accurately describe the system's heat transfer characteristics and nonlinear behavior. The system breaks down key components (such as heat exchangers, pumps, valves, and fans) into multiple dynamic partitions and defines state variables for energy accumulation and transfer in each partition, such as temperature, heat capacity, and heat transfer coefficient. These state variables are assigned corresponding time constants and update cycles to ensure effective control adjustments at different time scales. To address nonlinearity and inertia issues in the system, the system establishes various nonlinear models (such as temperature dependence and flow rate dependence) and dynamically adjusts the system's control parameters through recursive correction methods and real-time data identification to improve the consistency between the model and the actual operating state. Extended Kalman filtering (EKF) is used to stably estimate the system state and output reliable control signals even in the presence of noise and model errors.

[0079] The system utilizes a forward-looking decision-making mechanism (such as constrained model predictive control, MPC) to achieve real-time optimization. Through multi-level target layer weight orchestration and collaborative allocation strategies, the system can flexibly adjust the control strategy according to changes in operating conditions at different times. Increased wind speed will affect the wind-side heat exchange efficiency, and the system will adjust the penalty weight of the fan operation in a timely manner to reduce energy consumption. When the equipment approaches the operating boundary, the system will prioritize protecting hard constraints and adjust other soft constraints. Combined with an adaptive gain tuning mechanism and real-time system evaluation, the system can cope with complex changes in the operating environment, achieve efficient energy management, and quickly adjust the control strategy when equipment fails or operating conditions change to minimize energy consumption and ensure stable system operation.

[0080] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions will not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. An energy-saving coupled control method for a direct air-cooled assisted molten salt thermal storage system, characterized in that, Includes the following steps: S1. Defining and Establishing Constraints for Operating Condition Boundaries: Layer disturbance factors according to time scale and establish statistical characteristics, establish equipment operating condition boundary constraints that include hard and soft constraints, unify the energy balance caliber, conduct observability analysis of measurement points, and introduce virtual sensing. S2. Dynamic profiling parameter self-consistency: The system is divided into multiple dynamic partitions, a dynamic model containing nonlinearity and inertia is established, the model parameters are corrected online, and re-identification is triggered by monitoring the energy closure residual. S3, Multi-source data embellishment and reliable state reconstruction: Time alignment, delay compensation and anomaly weighting are performed on the measurement data, and the system state is estimated and the uncertainty is output using extended Kalman filtering; S4. Hierarchical weight arrangement of target layer construction: Construct a multi-objective function that includes energy consumption, tracking deviation, action frequency and boundary approximation penalty, perform dimensionless normalization and dynamically adjust weights, and set hard constraints to have absolute priority and soft constraints to be graded down. S5. Forward-looking decision kernel rolling solution: Rolling optimization is performed using a multi-rate model predictive control framework, combined with hot start and fault rollback mechanisms, to filter and freeze the changes of the online updated model parameters. S6. Adaptive identification gain tuning mechanism: Select key parameters through sensitivity analysis, identify them online using recursive least squares method with forgetting factor, map them to controller parameters, and set an abnormal freeze rollback mechanism. S7. Coupling Resource Orchestration and Cooperative Allocation: Introduces multi-loop coupling conflicts into constraint optimization, allocates resources based on the principle of minimum marginal cost, and uses consistency supervision quantities to suppress oscillations; S8. Execution layer decoupling-driven protection logic injection: Establish nonlinear mapping and decoupling protection logic for the execution mechanism, prioritize arbitration with hard constraints, and feed back the execution status to the estimator and optimizer; S9. Operational evaluation closed-loop knowledge accumulation and update: Hierarchical recording of operational data and events, root cause analysis, and version management and rollback of control strategies and parameters after playback verification.

2. The energy-saving coupling control method for a direct air-cooled assisted molten salt thermal storage system according to claim 1, characterized in that, Step S1 specifically includes: S1.

1. Ambient temperature, wind speed and direction, intraday irradiance variation, unit load fluctuation, and cold end available resource fluctuation are taken as disturbance factors and stratified according to time scales of seconds, minutes and hours, and defined as three categories: measurable, estimable and unobservable. At the same time, statistical characteristics including mean, variance, correlation time and typical transition amplitude are established. S1.

2. Take the heat exchange section, pump valve, bypass, fan / fin section, and heating section as key objects, and write the allowable working range of their material temperature limit, phase window, and pressure difference limit as hard constraints. Write the allowable working range of their temperature approaching the boundary and fan speed approaching the limit as soft constraints. Give the dimensions of each constraint, the consistency of the sampling period, and the out-of-bounds handling strategy including penalty, saturation, and degradation. S1.

3. Clarify the conventions for positive and negative heat values, flow rate benchmarks, enthalpy calculation benchmarks, rules for including pressure drop losses, and the placement of auxiliary power consumption in the energy balance, and solidify these into a data dictionary; S1.4 Map the layout of measurement points for temperature, differential pressure, flow rate, valve position, speed, and electrical power to a set of state variables, perform rank / sensitivity analysis, and if unobservable states or weakly controllable quantities are found, introduce alternative variables, virtual sensors, or supplementary measurement points at the strategy layer.

3. The energy-saving coupling control method for a direct air-cooled assisted molten salt thermal storage system according to claim 1, characterized in that, Step S2 specifically includes: S2.

1. Divide the system into several dynamic partitions. For each partition, select the partition average temperature, equivalent heat capacity, equivalent heat transfer coefficient, and flow hysteresis state as state variables. The selection principle is that the state energy can be closed and energy is conserved, it can be deduced from the measurement, and it is sensitive to the manipulated quantity. At the same time, the time constant and update cycle of each state are clearly defined. S2.2 The heat transfer coefficient changes with flow velocity, the physical properties change with temperature, the air-side heat transfer is coupled with the fan speed and the ambient wind speed, and the transport delay introduced by the pipeline / heat exchanger is treated as nonlinearity and inertia, and represented by differentiable functions. The pure delay is implemented using first-order / multi-order hysteresis or discrete shift registers. S2.3 First, use historical data to offline fit a set of initial values ​​for heat transfer / resistance / heat capacity under multiple representative working conditions. Then, during operation, use a recursive method to slowly correct the drift rate and add a regularization term, and set a confidence threshold: when sensor drift or anomaly occurs, pause the adaptive process and fall back to the conservative parameter set. S2.4 Calculate the energy closure residual for each sampling period, and allocate the residual to three channels according to the confidence level: measurement deviation, unmodeled thermal loss, and physical property error. Set up residual statistical monitoring. If the residual is biased for a long time, trigger the re-identification or measurement point calibration process.

4. The energy-saving coupling control method for a direct air-cooled assisted molten salt thermal storage system according to claim 1, characterized in that, Step S3 specifically includes: S3.

1. Perform timestamp correction and resampling to a uniform step size for each measurement point, and establish a channel delay model that includes sensor response, communication buffer, and filter group delay; perform delay compensation for key closed-loop quantities or use predicted values ​​to enter the controller. S3.

2. Identify jump points, deadlocks, drifts, and saturation as data ailments. Perform combined cleaning through physical boundary checks, differential mutation detection, consistency checks, and short-window statistical tests. Assign a confidence weight to each data point so that the estimator can automatically ignore suspicious observations. S3.

3. State estimation is performed using an extended Kalman filter under noise and model error conditions. Its discrete-time form is as follows: In the formula, Let be the system state vector. Its estimated value, To manipulate the input vector, To measure the output vector, For discrete state transition functions, For measurement function, Let Jacobian be the state transition with respect to the state. Let P be the Jacobian of the measurement function with respect to the state, Q be the process noise covariance, R be the measurement noise covariance, and K be the measurement noise covariance. k Kalman gain; S3.4 Output State Estimation The uncertainty P or its diagonal variance, including key derivatives such as heat flow, equivalent drag, and energy inventory, as well as the weights of each measurement point, are used by the controller for risk constraints or adaptive parameter tuning.

5. The energy-saving coupling control method for a direct air-cooled assisted molten salt thermal storage system according to claim 1, characterized in that, Step S4 specifically includes: S4.

1. Write the power consumption terms of the fan and pump as the convex penalty of the input, write the temperature / heat flow tracking as the deviation penalty, write the frequent actions as the input rate of change penalty, and write the approach to the boundary as the soft constraint relaxation penalty. S4.

2. Reduce the priority of wind turbine-related penalty weights when wind speed is high; increase the weight of motion smoothing terms when measurement point reliability is low; increase constraint relaxation penalties when equipment approaches the boundary. S4.3 Normalize each penalty to a dimensionless quantity using a reference value, which includes allowable deviation, rated power, and maximum rate of change, and record the source of each reference value; S4.

4. Set hard constraints to always take precedence; soft constraints are degraded in the order of releasing comfort zone type first, then efficiency type, and finally motion smoothness type, and define trigger conditions for each degrade level, including duration, out-of-bounds magnitude, and estimated uncertainty threshold.

6. The energy-saving coupling control method for a direct air-cooled assisted molten salt thermal storage system according to claim 1, characterized in that, Step S5 specifically includes: S5.1 Make the prediction time domain cover the main thermal inertia from minutes to hours, the control time domain shorter than the prediction time domain, and set up a multi-rate mechanism: fast loop handles the details of the actuator, and slow loop handles energy distribution and boundary management; S5.

2. Load the indicators and constraints defined in step S4 into the constraint model predictive control framework, using a quadratic form: In the formula, N p To predict the number of steps, N c To control the number of steps, ΔU is the control increment sequence. r k+i For reference trajectory, Q is the weighted L2 norm. y R u S and S are the weight matrices for output deviation, input increment, and soft constraint relaxation, respectively; A, B, and C are the discrete linear model matrices; d k+i This is the equivalent disturbance / modeling error compensation term. For the set of constraint functions that need to be softened, These are slack variables; S5.

3. Employ a mechanism that includes hot start, constraint projection, and reverting to the previous feasible control or switching conservative rules in case of failure, and bind the solver iteration count, termination condition, and sampling period. S5.

4. Apply amplitude limiting and low-pass filtering to the parameter changes obtained from online identification; set a freeze zone for changes in the key matrix, and do not update if the change is less than the threshold; and perform a feasibility check after each update to ensure that the constraints are met and the gain is normal.

7. The energy-saving coupling control method for a direct air-cooled assisted molten salt thermal storage system according to claim 1, characterized in that, Step S6 specifically includes: S6.1 First, perform sensitivity analysis, select the few parameters that have the greatest impact on key outputs to enter online updates, and reparameterize highly correlated parameters, merging two product terms into one equivalent parameter; S6.

2. The recursive least squares method with a forgetting factor is used to correct slowly changing equivalent parameters online. Its standard form is: In the formula, For the estimation of the parameter vector to be identified, In order to return to the goal, For the regression vector, The covariance is estimated for the parameters, where λ∈(0,1] is the forgetting factor. For RLS gain; S6.3, will The parameters are mapped to weights, model matrices, and disturbance compensation coefficients available within the controller for parameter tuning, and the physical rationality of the parameters, model stability and controllability are checked simultaneously, as well as the constraint boundaries are recalculated. S6.4 Set the freeze condition to sensor weights that are too low, energy residual statistics that are abnormal, or the solver that is frequently infeasible; hold after triggering. Remain unchanged and roll back to the most recent snapshot of health parameters, while also recording events.

8. The energy-saving coupling control method for a direct air-cooled assisted molten salt thermal storage system according to claim 1, characterized in that, Step S7 specifically includes: S7.

1. Write the coupling conflicts of multiple input pairs with multiple states and power consumption into the objectives and constraints in the form of cross terms, shared constraints, and coupled dynamics. S7.

2. Power consumption, near-boundary risk, and motion wear are converted into a unified marginal cost curve that is updated with the operating conditions, and the controller allocates power based on the principle of minimizing marginal cost. S7.

3. Represent the cooperative assignment as a constrained optimization model, using the augmented Lagrangian form: Its Lagrange function is: In the formula, u is the collaborative allocation decision variable. Let h(u) = 0 be the cost term of type j, h(u) = 0 be the equality constraint, g(u) ≤ 0 be the inequality constraint, c(u) be the coupling conflict quantity or residual term to be suppressed, λ and μ be the Lagrange multipliers, and ρ > 0 be the quadratic penalty coefficient. S7.4 Introduce a consistency supervision quantity. When oscillation signs are detected, increase the smoothing weight, shorten the control time domain, or increase damping, and limit the simultaneous large-scale movement of multiple channels.

9. The energy-saving coupling control method for a direct air-cooled assisted molten salt thermal storage system according to claim 1, characterized in that, Step S8 specifically includes: S8.1 Establish valve opening mapping, pump / fan speed / power mapping, and a limiting and dead zone processing model that includes minimum resolution, hysteresis, saturation, and start / stop constraints, and impose slope limits on commands. S8.2 Separate the master control quantity and slave quantity through feedforward cancellation, static decoupling matrix or cascade structure. When a conflict occurs, perform priority arbitration, satisfy hard constraint related loops first, and then process performance loops. S8.3 Make the trigger threshold, hold time, release condition, and regression slope of the protection logic consistent with the sampling period; when protection is triggered, simultaneously tighten the constraints of the upper-level optimizer or switch to the degradation target; S8.4 Real-time monitoring of command-feedback deviation, actuator jamming signs, saturation ratio and cumulative action cost, and writing back valid input feedback to the estimator and optimizer to replace command input for status updates.

10. The energy-saving coupling control method for a direct air-cooled assisted molten salt thermal storage system according to claim 1, characterized in that, Step S9 specifically includes: S9.1 Record key states, inputs, constraint occupancy rates, soft constraint relaxation amounts, estimation uncertainties, and solver states on time scales of seconds, minutes, and hours, and label abnormal, protection trigger, parameter freeze, and sensor calibration events. S9.2, Combining energy closure residuals and EKF innovative sequences Statistics and MPC infeasibility frequency are used to perform root cause decomposition of measurement bias, unmodeled loss, parameter drift, disturbance prediction error, and actuator failure, and provide processing actions such as calibration, re-identification, adjustment of noise covariance, or tightening of constraints. S9.

3. Perform replay verification on typical scenarios including high disturbance, low load, boundary approach, and sensor anomaly to check whether infeasibility occurs, whether too many protections are triggered, or whether input oscillation occurs. Only after passing the verification should the new weight rules, new parameter initial values, and new thresholds be written to the version library. S9.4 records the version number, change item, applicable working conditions, and rollback path for each change, and incorporates key hyperparameters Q, R, λ, weight matrix, and soft constraint penalties into configuration management, allowing the runtime to roll back to the previous stable version with one click.

Images