Industrial park-oriented carbon energy intelligent control system

Abstract

The application discloses an industrial park-oriented carbon control system, relates to the field of computer technology and intelligent energy management, and comprises the following: an instruction receiving module receives a load reduction instruction issued by a power grid dispatching center; a state collecting module collects real-time physical state and environmental temperature data of target materials after cutting off power supply of a target pipeline; a phase change evaluation module inputs the data into a pre-trained phase change restart cost prediction model to output real-time crust thickness and marginal restart cost; a benefit prediction module inputs real-time electricity price and carbon trading price into a pre-trained deep Q network to output marginal trading benefit; and an instruction executing module compares the values of the marginal restart cost and the marginal trading benefit, generates a power supply recovery instruction and recovers power supply of the target pipeline when the marginal restart cost is equal to the marginal trading benefit. The application solves the problem that the existing dispatching system ignores the phase change characteristics of materials, which leads to an exponential increase in restart energy consumption after a long shutdown and thus seriously counteracts the trading benefit.

Description

Technical Field

[0001] This invention relates to the fields of computer technology and smart energy management, and more specifically, to an energy and carbon intelligent control system for industrial parks. Background Technology

[0002] Industrial parks, as hubs of modern industrial production, are core contributors to energy consumption and carbon emissions. With the continuous improvement of electricity market trading mechanisms, parks are no longer merely simple energy-consuming terminals, but rather important regulatory resources capable of participating in grid demand response and high-price power dispatch through intelligent energy and carbon control systems. In the existing technological landscape, Chinese patent document CN109523052A discloses a virtual power plant optimal dispatch method considering demand response and carbon trading. This technology establishes an optimal dispatch model by statistically analyzing the number of controllable loads within the virtual power plant and based on day-ahead forecast information and price information from the electricity and carbon trading markets, thereby leveraging the advantages of demand response resources in peak shaving and valley filling. Meanwhile, Chinese patent document CN116596123A discloses a low-carbon optimal dispatch method for integrated energy systems in industrial parks. This scheme establishes a tiered carbon trading mechanism model for industrial parks with compensation coefficients and performs multi-objective optimization calculations with the minimum operating cost, abandoned power, and carbon emissions as objective functions.

[0003] However, existing systems generally suffer from a technological blind spot when executing shutdowns for power generation or carbon quota sales during periods of high electricity prices, as they are detached from the physical realities of industrial production. Current technologies typically focus only on macroscopic electricity price differences and the book profits from carbon trading, treating the scheduled industrial production lines as idealized digital nodes that can be shut down and restored at any time, completely ignoring the physical evolution of intermediate materials within complex industrial pipelines during power outages and cooling periods. For materials involving molten metals or temperature-controlled reaction liquids, which are extremely sensitive to temperature, prolonged power outages cause the materials to cool and solidify, forming a gradually thickening solid crust on the pipeline's inner wall. This non-linear phase change process drastically reduces the effective flow cross-sectional area of ​​the pipeline. When the system ends the scheduling and restarts the production line, an exponentially increasing amount of thermal energy is required to remelt the crust, while the sudden change in fluid resistance leads to severe overload of pumping power. This hidden and costly restart energy penalty often far outweighs the short-term book profits from selling electricity during shutdowns. If an extremely conservative fixed-time countdown is used to prevent solidification, it is easy to miss out on a large amount of grid trading benefits. If the power supply time is blindly pursued, resulting in deep solidification of the pipeline, it will eventually fall into a negative benefit dilemma of earning meager power supply fees but paying exorbitant restart costs. How to balance the two is a problem that needs to be solved in the existing technology. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to provide an energy and carbon intelligent control system for industrial parks, so as to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: An energy and carbon intelligent control system for industrial parks includes an instruction receiving module, a status acquisition module, a phase change assessment module, a revenue prediction module, and an instruction execution module. The instruction receiving module receives load reduction instructions issued by the power grid dispatch center; Upon receiving the load reduction command, the status acquisition module cuts off the power supply to the target pipeline and collects real-time physical status data of the target material and ambient temperature data within the target pipeline through sensors. The phase change assessment module inputs the real-time physical state data and the ambient temperature data into the pre-trained phase change restart cost prediction model, and the pre-trained phase change restart cost prediction model outputs the real-time crust thickness and marginal restart cost of the inner wall of the target pipeline. The revenue prediction module acquires real-time electricity price data and real-time carbon trading price data, and inputs the real-time electricity price data and the real-time carbon trading price data into a pre-trained deep Q network. The pre-trained deep Q network outputs the marginal trading revenue for continuing to execute the load reduction command. The instruction execution module compares the marginal restart cost with the marginal transaction benefit. When the marginal restart cost equals the marginal transaction benefit, it generates a power restoration instruction and restores power to the target pipeline.

[0006] Specifically, the phase change restart cost prediction model includes a feature extraction network, a shell thickness prediction branch network, and a restart cost evaluation branch network; The feature extraction network receives the real-time physical state data and the ambient temperature data as input, and outputs a hidden layer feature vector. The crust thickness prediction branch network receives the hidden layer feature vector as input and outputs the real-time crust thickness of the inner wall of the target pipeline. The restart cost evaluation branch network receives the hidden layer feature vector and the real-time shell thickness as input, and outputs the marginal restart cost.

[0007] Specifically, the training process of the phase change restart cost prediction model includes: Acquire historical physical state data, historical ambient temperature data, historical actual crust thickness data, and historical actual restart power consumption data; The historical physical state data and the historical ambient temperature data are input into the phase change restart cost prediction model to be trained, and the predicted crust thickness and the predicted marginal restart cost are output. The first loss value is obtained by subtracting the predicted crust thickness from the historical actual crust thickness data. The second loss value is obtained by subtracting the predicted marginal restart cost from the historical actual restart power consumption data. The first loss value and the second loss value are summed to obtain the total loss value. The network parameters of the phase transition restart cost prediction model to be trained are updated using the backpropagation algorithm until the total loss value converges to below the convergence threshold, thus obtaining the trained phase transition restart cost prediction model.

[0008] Specifically, the real-time physical state data includes the dynamic viscosity data and real-time core temperature data of the target material; The marginal restart cost is the sum of the heat energy consumption for crust melting and the additional power consumption for pumping required to terminate the load reduction command and start the main material pump connected to the target pipeline at the current moment. The instruction execution module is equipped with a difference comparator. The difference comparator calculates the difference between the marginal transaction revenue and the marginal restart cost. When the difference shows a decreasing trend and equals zero within three consecutive time steps, the operation of generating the power restoration instruction is triggered.

[0009] Specifically, the energy and carbon intelligent control system for industrial parks also includes a micro-circulation frequency modulation module; The micro-circulation frequency modulation module maintains the power supply to the material circulation pump connected to the target pipeline during the power outage of the target pipeline; The micro-circulation frequency modulation module inputs the real-time shell thickness and the preset safe shell thickness threshold into the reinforcement learning hovering control model, and the reinforcement learning hovering control model outputs the target operating frequency. The micro-circulation frequency modulation module adjusts the operating frequency of the material circulation pump to the target operating frequency.

[0010] Specifically, the architecture of the reinforcement learning hovering control model adopts a deep deterministic policy gradient architecture, which includes an actor network and a critic network. The state space elements of the reinforcement learning hovering control model include the real-time crust thickness, the preset safe crust thickness threshold, and the operating frequency of the material circulation pump at the previous moment. The action space elements of the reinforcement learning hovering control model are the target running frequencies; The actor network outputs the target running frequency, and the critic network outputs the state-action value score corresponding to the target running frequency.

[0011] Specifically, the reinforcement learning hovering control model uses an environmental reward function to update parameters during the training phase; The environmental reward function includes a thickness maintenance reward term and an energy consumption penalty term; When the next moment is reached, if the absolute value of the difference between the real-time shell thickness obtained by the phase transition evaluation module and the preset safe shell thickness threshold is less than the preset difference threshold, the thickness maintenance reward item outputs a positive reward. The value of the energy consumption penalty term is inversely proportional to the value of the target operating frequency; The environmental reward function sums the thickness maintenance reward term and the energy consumption penalty term to obtain a total reward value, and uses the total reward value to update the weight parameters of the actor network and the critic network.

[0012] Specifically, the energy and carbon intelligent control system for industrial parks also includes a gas storage pre-compression module and a pulse ice-breaking module; During the time period before the command receiving module receives the load reduction command issued by the power grid dispatch center, the gas storage pre-compression module sends a start signal to the air compressor connected to the gas pressure pipeline network, and raises the gas pressure of the gas storage tank connected to the air compressor to the upper limit of the safe pressure withstand threshold of the gas pressure pipeline network, and then cuts off the power supply to the air compressor. The pulse ice-breaking module blocks the starting power circuit of the main material pump connected to the target pipeline when the instruction execution module generates the power restoration instruction, and at the same time sends an opening signal to the pneumatic solenoid valve connected in parallel with the target pipeline.

[0013] Specifically, the pulse ice-breaking module inputs the real-time crust thickness into the pneumatic pulse parameter matching template library; The pneumatic pulse parameter matching template library stores a mapping table between different shell thicknesses and corresponding opening durations. The pulse ice-breaking module retrieves the target activation duration from the pneumatic pulse parameter matching template library; After receiving an opening signal, the pneumatic solenoid valve remains open to release the high-pressure gas in the gas storage tank into the target pipeline, and closes after the target opening duration has elapsed.

[0014] Specifically, the instruction execution module is equipped with a time-delay relay; The input terminal of the time delay relay receives the power restoration command generated by the command execution module, and the output terminal of the time delay relay is connected to the start control terminal of the main material pump; The time delay relay has a time delay parameter inside, and the value of the time delay parameter is equal to the target activation duration plus a preset oscillation calming duration; After receiving the power restoration command for the specified time period, the time delay relay activates the starting power circuit of the main material pump.

[0015] The advantage of this invention over existing technologies lies in its deep integration of fluid thermodynamics phase transition laws and the game theory logic of energy trading market economics. It cleverly solves the pain point of existing scheduling systems where restart energy consumption penalties negatively impact scheduling benefits due to detachment from material physical properties. After acquiring real-time pipeline physical state data and ambient temperature data, this invention inputs them into a pre-trained phase transition restart cost prediction model to accurately calculate the real-time crust thickness of the pipeline's inner wall and the marginal restart cost. Simultaneously, external market price data is input into a pre-trained deep Q-network to output marginal trading revenue. By comparing these two values, the system accurately issues a power restoration command at the critical moment when the marginal restart cost equals the marginal trading revenue. This design proactively allows for moderate crust formation on the pipeline's inner wall, utilizing the latent heat released during the crusting process and the insulation properties of the solid shell to maintain the temperature of the central material. This transforms the originally dangerous physical phase transition process into a controllable insulation barrier, ensuring the safety of industrial pipelines during power outages from a fundamental logical perspective while maximizing the effective outage benefit period for participating in grid peak shaving.

[0016] Furthermore, this invention addresses the problem of excessively rapid crust formation when materials are stationary and power is off, leading to insufficient downtime for high-value operations. By introducing a micro-circulation frequency modulation module and a reinforcement learning hovering control model, the system maintains the weak operation of the material circulation pump with extremely low power consumption during power outages, utilizing the shear friction self-heating effect of the fluid within the narrowed pipe to offset heat loss. This ingenious application of physical viscous frictional heat forcibly keeps the crust thickness hovering near a safe threshold, significantly extending the window for generating profits during power outages with extremely low additional energy consumption.

[0017] To address the issue of massive power spikes and subsequent grid demand penalties that can easily occur when power must be restored and restarted near the phase transition critical point, this invention creatively resolves this problem through a gas storage pre-compression module and a pulse ice-breaking module. During periods of grid parity, the system pre-fills the main gas pressure pipeline with high-pressure gas. At the moment when a heavily loaded pipeline must restart, the system prioritizes releasing a high-pressure pneumatic pulse into the pipeline, utilizing water hammer acoustic wave oscillations to internally disrupt the static frictional engagement of the solid crust, followed by a time delay before powering on the main pump. This mechanism successfully converts the pre-stored pneumatic potential energy into mechanical ice-breaking kinetic energy, achieving a smooth, zero-surge restart of high-resistance pipelines. Attached Figure Description

[0018] Figure 1 shows a physical diagram of the overall architecture of the present invention, including the physical building exterior representing the power grid dispatch center, the industrial control cabinet hardware with indicator lights representing the energy and carbon intelligent control system, and physical pipelines with flanges, and shows the actual physical cable connections.

[0019] Figure 2 is a schematic cross-sectional view of the target pipeline cross-section state acquisition of the present invention, showing the concentric circles of the internal material and the crust layer, the metal flange and fastening bolt hole structure of the pipeline outer wall, and a solid needle-shaped sensor probe that penetrates the pipe wall and the crust layer and directly pierces into the liquid phase material in one embodiment.

[0020] Figure 3 illustrates the integrated circuit board diagram of the phase change restart cost prediction model of the present invention, including a printed circuit board with pads, in which the feature extraction network is presented as the main central processing unit, and the two branch networks are presented as coprocessor chips.

[0021] Figure 4 is a schematic diagram of the hardware structure of the instruction execution module logic comparison. The comparison logic is converted into the actual operational amplifier hardware circuit symbol in electronic engineering. Its output is directly connected to the timing record control integrated chip with hardware clock identification.

[0022] Figure 5 This diagram illustrates the reinforcement learning hovering control industrial programmable logic controller of the present invention. The algorithm model is transformed into a schematic diagram of a PLC rack backplane commonly found in industrial settings. The actor network and the critic network are drawn as independent logic processing hardware modules plugged into the backplane, with terminal blocks. Detailed Implementation

[0023] The specific embodiments of the present invention will now be described with reference to the accompanying drawings.

[0024] This invention is applicable to pipelines within industrial parks that transport high-temperature molten materials, constant-temperature reaction liquids, high-viscosity fluids, or other materials prone to cooling and crusting during power outages. Typical scenarios include molten salt transport pipelines, molten metal insulation pipelines, asphalt transport pipelines, and high-temperature chemical mother liquor circulation pipelines. For these production lines, power restoration during shutdown is not simply a matter of cutting off a single motor; it triggers a chain reaction of material temperature drop, viscosity increase, localized solidification, reduced flow cross-section, and a sudden increase in restart resistance. This invention establishes control logic around this physical process, enabling industrial parks to participate in grid demand response, electricity price arbitrage, and carbon quota trading by simultaneously assessing the phase change costs occurring within the pipeline after shutdown, rather than just looking at paper profits. This allows for a comparison of shutdown benefits and restart costs on the same decision-making plane.

[0025] When industrial parks face exceptionally high summer electricity loads or near-depletion of carbon emission allowances, the instruction receiving module prioritizes listening for and receiving load reduction instructions from the power grid dispatch center. These instructions typically include substantial electricity price compensation to encourage businesses to proactively cut off non-emergency power consumption. In one embodiment, the energy and carbon intelligent control system for industrial parks is deployed across the park's energy management master station, edge controllers, and field execution units, such as... Figure 1As shown, the energy management master station is responsible for receiving instructions from the power grid dispatch center, calling the prediction model, and making revenue decisions. The edge controller is responsible for sensor acquisition, equipment interlocking, and field execution. The field execution units include circuit breakers, frequency converters, relays, pneumatic solenoid valves, air tanks, air compressors, and various sensors. To ensure continuous acquisition of critical status during power outages, the status acquisition link, edge controller, and critical valve control power supply are preferably connected to the uninterruptible power supply branch, while the main heating branch and the high-energy-consuming main pump branch are connected to the disconnectable load branch. The purpose of this setup is to isolate the high-energy-consuming equipment that actually participates in power trading from the low-energy equipment that maintains safety awareness and minimum control capabilities, avoiding the system entering a blind state after a sudden power outage.

[0026] In one embodiment, the system includes an instruction receiving module, a status acquisition module, a phase change evaluation module, a revenue prediction module, and an instruction execution module. The instruction receiving module is connected to the park dispatch communication gateway and is used to receive load reduction instructions issued by the power grid dispatch center. This load reduction instruction can carry the load reduction start time, target reduction power, estimated duration, response level, and permissible recovery conditions. After the instruction arrives, the instruction receiving module first performs legality verification, timestamp verification, and equipment availability verification, and then sends the confirmed dispatch parameters to the status acquisition module and the instruction execution module. Verification before execution avoids communication jitter, duplicate or expired instructions directly entering the field layer, thus preventing unnecessary downtime.

[0027] In this embodiment, after receiving a load reduction command, the status acquisition module cuts off the main power supply circuit corresponding to the target pipeline. The preferred targets for this cutoff are the main heating equipment, insulated electric heat tracing branch, main material pump, and other high-energy-consuming auxiliary equipment directly related to the target pipeline, while the status acquisition link, communication link, and safety interlock power supply continue to be powered. For the material circulation pump involved in the micro-circulation frequency modulation module described later, an independent controlled branch power supply can be used, allowing it to continue operating at low frequency according to the strategy even after the main branch is disconnected. This approach is because the present invention aims for a bounded, flexible shutdown, rather than a hard shutdown that completely loses its regulatory capability.

[0028] The status acquisition module collects real-time physical state data of the target material within the target pipeline and ambient temperature data via sensors. The real-time physical state data includes at least dynamic viscosity data and real-time core temperature data. Dynamic viscosity data can be provided by an online rotary viscometer, a vibratory viscometer, or a combined inversion module based on differential pressure and flow rate. Real-time core temperature data can be acquired by an insertion thermocouple, a high-temperature resistant fiber optic temperature probe, or a multi-point temperature array arranged along the pipe diameter. Ambient temperature data is preferably collected by a pipe rack ambient temperature sensor, an equipment floor space temperature sensor, or an outdoor weather station. The sampling period can be set from 1 second to 60 seconds, with a suitable range of 5 to 10 seconds. A sampling period that is too long will result in untimely capture of crust evolution, while a period that is too short will introduce unnecessary communication and computational burdens. To suppress single-point noise, the status acquisition module can also perform a moving average of 3 to 5 consecutive sampled values ​​and remove anomalies from data exceeding the material process boundaries. In some embodiments, such as... Figure 2 As shown, this invention employs a solid needle-shaped sensor probe that penetrates the pipe wall and crust layer, directly piercing deep into the liquid material. The direct-penetration probe structure, coupled with a high-temperature, high-pressure resistant titanium alloy protective sleeve, avoids interference from the material layer that solidifies first at the pipe wall. The probe tip precisely integrates a miniature vibrating wire viscometer and a high-sensitivity thermocouple temperature sensing node. The vibrating wire maintains a specific frequency of micro-oscillation in the fluid; by measuring the attenuation of the amplitude, the dynamic viscosity data of the target material can be accurately obtained. The thermocouple is responsible for acquiring real-time core temperature data. These two data points constitute the real-time physical state data upon which all subsequent intelligent algorithms rely. The physical reason for introducing dynamic viscosity monitoring is that the viscosity of many polymeric fluids or molten metals undergoes a non-linear, drastic change as they approach their freezing point. This is the most crucial and proactive physical characteristic for predicting impending complete pipeline blockage.

[0029] In this invention, the phase change assessment module inputs real-time physical state data and ambient temperature data into a pre-trained phase change restart cost prediction model. This model outputs the real-time crust thickness of the target pipeline's inner wall and the marginal restart cost. The introduction of the real-time crust thickness as an intermediate variable is not to increase the computational hierarchy, but to break down the originally purely black-box restart energy consumption prediction into two segments that are more in line with physical intuition. Technicians can use the crust thickness to see whether the pipeline is currently in a stage of slight wall adhesion, localized annular narrowing, or deep solidification, thus understanding why the restart cost increases sharply for the same additional 5 minutes of downtime. This interpretability of the intermediate physical quantity significantly improves the system's acceptability in industrial settings.

[0030] In one embodiment, the phase change restart cost prediction model includes a feature extraction network, a crust thickness prediction branch network, and a restart cost evaluation branch network, such as... Figure 3As shown, the feature extraction network receives real-time physical state data and ambient temperature data as input and outputs hidden layer feature vectors. Considering the significant temporal sequence of viscosity and temperature changes over time, the feature extraction network preferably adopts a time-series modeling structure, such as a combination of a 1D convolutional network and gated recurrent units, or a combination of a multi-layer fully connected network and long short-term memory units. The front-end convolutional part is used to capture the temperature drop rate, viscosity rise slope, and environmental disturbance patterns, while the recurrent units are used to retain the historical effects of the previous few time steps, because the crust formation is not determined solely by the current instantaneous temperature, but is the result of the combined effects of continuous cooling history after shutdown. The dimension of the hidden layer feature vector can be set from 32 to 256 dimensions, with 64 or 128 dimensions being more suitable, as they are easier to balance between expressive power and deployment computing power.

[0031] The crust thickness prediction branch network receives the hidden layer feature vector as input and outputs the real-time crust thickness of the target pipeline's inner wall. This branch can adopt a 2- to 4-layer fully connected structure, and the output value can be either a direct millimeter value or a dimensionless proportional value relative to the pipeline's inner radius. Using millimeter values ​​makes it easier for field personnel to understand and compare with maintenance experience; using proportional values ​​is more suitable for model transfer between pipelines of different diameters. The real-time crust thickness does not require verification through physical disassembly and inspection every second; it is essentially an online estimate, but this estimate must be consistent with historical calibration data to reflect both trends and absolute levels.

[0032] The restart cost assessment branch network receives the hidden layer feature vector and real-time crust thickness as input, and outputs the marginal restart cost. Inputting the real-time crust thickness back into the cost branch explicitly allows the cost network to use the key physical intermediary of crust formation, preventing the model from inferring costs indirectly from implicit representations of viscosity and temperature. The marginal restart cost can be understood as the additional cost incurred when terminating the load reduction command and starting the main material pump connected to the target pipeline at the current moment to complete crust melting, restore flow path patency, and overcome additional flow resistance. This cost comprises two parts: the thermal energy consumption for crust melting and the additional power consumption for pumping. The former corresponds to the power consumed in reheating the solid crust back to a flowable state, while the latter corresponds to the additional power consumed by the main pump to restore flow after pipe diameter reduction, increased viscosity, and increased static friction. In actual deployment, the marginal restart cost is preferably uniformly converted into a monetary value, maintaining the same dimension as the marginal transaction revenue for easy direct comparison. If the training labels mainly collect additional power consumption data during restarts, the control platform can perform cost conversion based on the current electricity price during the power restoration period after the model outputs. If the park already has complete energy price labels, the conversion layer can also be incorporated into the output to directly output the cost value.

[0033] The training process for the phase change restart cost prediction model can be implemented as follows: First, acquire historical physical state data, historical ambient temperature data, historical actual crust thickness data, and historical actual restart power consumption data. Historical physical state data and historical ambient temperature data can come from long-term operational archives of the industrial park or from specially organized shutdown tests. Historical actual crust thickness data can be obtained through ultrasonic thickness measurement, acoustic emission inversion, fiber optic array temperature field inversion, post-shutdown disassembly and inspection section measurement, and joint calibration with thermal simulation. The reason for using joint calibration is that the crust state inside industrial pipelines is difficult to observe directly like laboratory samples; often, it is necessary to combine locally measurable results with thermodynamic models to form a reliable label. Historical actual restart power consumption data is preferably collected jointly from the main heating equipment meters, main pump meters, and auxiliary equipment meters during the recovery phase. The statistical window can be set to the entire process from the issuance of the recovery command to the stabilization of flow rate and center temperature, for example, 1 to 30 minutes.

[0034] During training, historical physical state data and historical ambient temperature data are input into the phase change restart cost prediction model to be trained, outputting the predicted crust thickness and the predicted restart cost. In one implementation, if the second supervision label uses historical real restart power consumption data, the predicted restart cost is preferably initially represented as the predicted restart power consumption, and then converted into the marginal restart cost in the deployment phase by combining the corresponding electricity price. In another implementation, the historical real restart power consumption data can also be first converted into historical real marginal restart cost data by combining the corresponding electricity price during the power restoration period, and then the model directly outputs the predicted marginal restart cost. For ease of explanation, the following example uses the model directly outputting the marginal restart cost. The first loss value is obtained by subtracting the predicted crust thickness from the historical real crust thickness data, and the second loss value is obtained by subtracting the predicted marginal restart cost from the historical real marginal restart cost data. The first loss value and the second loss value are then summed to obtain the total loss value. The network parameters are updated using the backpropagation algorithm until the total loss value converges to below the convergence threshold, resulting in the trained phase change restart cost prediction model. To avoid one branch dominating training due to different units of measurement, it is preferable to normalize the crust thickness label and restart cost label before actual training. The convergence threshold can be set to a normalization error range of 0.0001 to 0.01 according to the requirements of the site. In industrial settings, a range of 0.001 to 0.005 is more commonly used, as it is easier to achieve a balance between training stability and modeling accuracy. The optimizer can adopt an adaptive gradient optimization method, with a learning rate preferably set to 0.00001 to 0.001 and a batch size preferably set to 32 to 256. The purpose of this setting is not to pursue the ultimate accuracy in academic studies, but to enable the model to be stably deployed on an edge industrial control computer and to have the ability to be retrained when material formulations, seasonal environments, and process conditions change.

[0035] In this invention, the revenue prediction module acquires real-time electricity price data and real-time carbon trading price data, and inputs these data into a pre-trained deep Q-network. The deep Q-network then outputs the marginal trading revenue for continuing to execute load reduction orders. A deep Q-network is used here because the revenue of the industrial park participating in the transaction depends not only on the absolute level of the electricity or carbon price at a given moment, but also on whether it can maintain its eligibility to respond after a certain time step of continued shutdown, whether it remains in a high-price range, whether the price subsequently falls, and whether the timing of selling carbon allowances is still advantageous. In other words, whether to continue the shutdown or resume immediately is essentially a time-series decision problem with a cumulative effect over time, and deep Q-networks are more suitable for handling this type of strategy revenue assessment.

[0036] In one embodiment, the deep Q-network receives at least real-time electricity price data and real-time carbon trading price data as input. To improve the stability of the results, it can further concatenate price sequences from the most recent time steps, the duration of executed outages, the current allowable load reduction margin, the park's response completion rate, and the crusting risk level given by the phase change assessment module. It is emphasized here that price data and carbon price data are the core inputs in this invention, while other quantities can be added as enhancing features. The network architecture can adopt a 3- to 5-layer fully connected network, and the number of hidden units can be set to 64 to 512. If the on-site price fluctuations have strong short-term memory characteristics, a time series encoding layer can also be added at the input end. In one implementation, the network output can directly provide the marginal trading revenue corresponding to continuing the load reduction action; in another implementation, it can first output the action values ​​of continuing reduction and immediate recovery, and then the post-processing unit takes the difference between the two and maps it to the marginal trading revenue of continuing the load reduction instruction. The former form is easier to deploy directly, while the latter form facilitates the subsequent strategy extension to multi-action scenarios, such as layered recovery, partial recovery, and continued suspension.

[0037] The training of deep Q-networks can rely on historical electricity price curves, historical carbon trading price curves, historical demand response settlement data of industrial parks, and historical execution records to establish a simulation environment. The reward value for each time step is preferably defined as the net revenue generated by continuing load reduction at that time step. This net revenue can include electricity sales revenue, demand response compensation revenue, and carbon quota trading revenue, but can also deduct default risk reserves, execution deviation penalties, and equipment switching loss reserves. The network trained in this way outputs not static book revenue, but rather the marginal strategy value incorporating subsequent trends. During training, an experience replay mechanism and a target network mechanism can be used to reduce value estimation oscillations. The discount factor is preferably set to 0.85 to 0.99, and the initial exploration probability is preferably set to 0.05 to 0.2, gradually decreasing as training progresses. The experience pool capacity can be set to 10,000 to 1,000,000 records. This design is because market prices and industrial park execution scenarios vary significantly; without experience replay and a target network, the model is easily swayed by local market conditions and cannot form a stable decision boundary.

[0038] The instruction execution module compares the marginal restart cost with the marginal transaction benefit. When the marginal restart cost equals the marginal transaction benefit, it generates a power restoration command and restores power to the target pipeline. The physical meaning here is clear: before this critical point, the market benefit of an additional time step of downtime still outweighs the future restart cost, making continued downtime worthwhile; after this critical point, the additional benefit from an additional time step of downtime is insufficient to cover the additional cost of shell formation, and further downtime will lead from the profitable zone into the loss zone. Therefore, this invention does not aim for absolute shell prevention, but rather allows for moderate shell formation within a safe boundary, utilizing the latent heat released during shell formation and the insulating effect of the formed solid shell to extend the tradable downtime window.

[0039] In one embodiment, the instruction execution module includes a difference comparator, such as... Figure 4As shown. The difference comparator calculates the difference between the marginal transaction revenue and the marginal restart cost in real time. When the difference shows a decreasing trend over three consecutive time steps and equals 0, the operation of generating a power restoration command is triggered. Setting a decreasing trend over three consecutive time steps aims to filter out misjudgments caused by instantaneous market price spikes, sensor noise, and model output jitter, and to avoid repeated switching due to occasional intersections at a single sampling point. In engineering implementation, a strictly zero value rarely occurs naturally. Therefore, a zero-bandwidth judgment method can be used, considering a difference falling within a preset zero threshold band as equal to 0. This zero threshold band can be set to 0.5% to 5% of the current marginal transaction revenue, or to RMB 1 to RMB 100 in absolute currency values. The specific value can be determined based on the park's energy scale and transaction granularity. The time step is preferably consistent with the state sampling period, or it can be an integer multiple thereof. For example, when the sampling period is 5 seconds, the comparison period can be set to 5 seconds or 10 seconds. This setting ensures that the evaluation link and the execution link have consistent timing.

[0040] In the specific execution of power restoration, the command execution module can operate according to a tiered restoration strategy. For scenarios with only basic functions, the power restoration command preferentially restores the heat tracing or heating branch first, and then restores the main material pump branch when the preset start-up conditions are met, so that the material receives a certain amount of thermal compensation before being propelled to flow. For scenarios with a pulse ice-breaking module, the power restoration command does not immediately turn on the main material pump, but instead first enters the pneumatic ice-breaking and time-delay interlocking process, which will be explained in detail later. This design is because the most dangerous stage of crust formation is not the power outage itself, but the instant of restart. If the main pump is hard-started under high resistance, it is very easy to generate current peaks and pipeline impacts.

[0041] In a further embodiment, the system also includes a micro-circulation frequency modulation module. This module maintains power supply to the material circulation pump connected to the target pipeline during the interruption of the main power supply circuit of the target pipeline. Maintaining power supply here does not mean running the circulation pump at the normal process frequency, but rather preserving a weak circulation capability only under the premise of extremely low power consumption. The reason for introducing micro-circulation is that when the material is completely still, the area near the pipe wall cools down first, and crust formation rapidly develops from the pipe wall inwards. However, once controlled weak flow exists, the fluid generates shearing and viscous frictional heat within the narrowed flow channel. Although this self-heating is not large, it is sufficient to significantly slow down the rate of crust thickening in many material systems. In other words, this invention does not simply rely on external heating to prevent solidification, but rather utilizes the fluid's own flow to generate some energy internally, thereby extending the trading window with lower energy consumption.

[0042] The micro-circulation frequency modulation module inputs the real-time crust thickness and the preset safe crust thickness threshold into the reinforcement learning hovering control model. The model then outputs the target operating frequency, and adjusts the operating frequency of the material circulation pump to the target frequency. Hovering here can be understood as controlling the crust thickness within a small range of fluctuations around the safe threshold, preventing it from continuously and rapidly thickening while also preventing the circulation pump from running at high frequencies for extended periods, thus avoiding excessive energy consumption. In other words, the system does not attempt to completely eliminate the crust, as that would consume a lot of extra electricity, which is not in line with the economic goals of an energy-saving scenario; the system truly aims for stable operation near the safe boundary.

[0043] In one embodiment, the reinforcement learning hovering control model employs a deep deterministic policy gradient architecture, such as... Figure 5 As shown, the architecture comprises an actor network and a critic network. The actor network outputs the target operating frequency, and the critic network outputs the state-action value score corresponding to the target operating frequency. Since the operating frequency is a continuous value, a deep deterministic strategy gradient architecture is more suitable than a discrete action strategy, as it can directly obtain a smooth frequency control quantity, avoiding the frequency from only being able to jump between a few fixed levels, which would cause significant fluctuations in crust thickness and pump energy consumption. In this embodiment, the state space elements of the reinforcement learning hovering control model include the real-time crust thickness, the preset safe crust thickness threshold, and the operating frequency of the material circulation pump at the previous moment. This setting corresponds to the current risk, the target boundary, and the control inertia, respectively. The current risk tells the model how far the crust has developed, the target boundary tells the model around which level to adjust, and the operating frequency at the previous moment helps the model realize that the pump speed cannot be changed instantaneously without cost, and control continuity must be considered. Without changing the core idea of ​​this invention, the state space can also be enhanced with variables such as real-time center temperature, dynamic viscosity, and ambient temperature to further improve adaptability under different seasons and processes. The action space element is the target operating frequency. The target operating frequency is preferably set within the range of 5% to 40% of the rated frequency, with a more suitable range of 10% to 30%. If converted to a power frequency of 50 Hz, the commonly usable range is 5 Hz to 20 Hz. This range provides a certain amount of shear self-heating without pulling the cycle energy consumption back to near normal production levels.

[0044] The preset safety crust thickness threshold can be determined based on a combination of pipe diameter, material, target flow rate, and allowable pressure drop. For metal insulated pipelines, the safety crust thickness threshold can be set to 5% to 25% of the pipeline's inner radius, or directly to 1 mm to 20 mm. The smaller the pipe diameter, the higher the material viscosity, and the smaller the main pump head margin, the more conservative the safety threshold should be set. To provide a clear reward criterion for hover control, the preset difference threshold can be set to 1% to 10% of the safety crust thickness threshold, or to 0.1 mm to 2 mm. A difference threshold that is too small will cause the model to overemphasize line-fitting control, making the frequency output prone to jitter; a difference threshold that is too large will make the control too coarse, losing the meaning of hovering.

[0045] The reinforcement learning hovering control model uses an environmental reward function for parameter updates during the training phase. This function includes a thickness maintenance reward term and an energy consumption penalty term. At the next time step, when the absolute value of the difference between the real-time shell thickness obtained by the phase transition evaluation module and the preset safe shell thickness threshold is less than the preset difference threshold, the thickness maintenance reward term outputs a positive reward. This positive reward essentially tells the model that the current control action has maintained the shell near an acceptable safe boundary, and fine-tuning can continue along this control direction. The energy consumption penalty term is introduced as a negative reward term into the total reward. Its absolute value is preferably increased with the target operating frequency, reflecting that a higher operating frequency corresponds to a higher energy consumption burden. The total reward value is obtained by summing the thickness maintenance reward term and the energy consumption penalty term, and then using this total reward value to update the weight parameters of the actor network and the critic network. This approach aims to allow the model to simultaneously consider shell suppression effects and micro-circulation energy consumption, avoiding the need to maintain excessively high operating frequencies for extended periods in pursuit of thickness stability.

[0046] For the training platform, the reinforcement learning hovering control model is preferably trained offline in a digital twin environment or a historical operating condition playback environment. The digital twin environment can reproduce the crust evolution process under different ambient temperatures, material formulations, initial crust thicknesses, and pipe diameters. The learning rate of the actor network and the critic network can be set to 0.00001 to 0.001, the discount factor to 0.9 to 0.99, the soft update coefficient to 0.001 to 0.02, and the experience pool capacity to 100,000 to 1,000,000 records. After training, the model is then embedded into the park's edge controller for online operation. To further ensure on-site stability, the micro-circulation frequency modulation module can also set an upper limit on the rate of change of the target operating frequency between two adjacent time steps, for example, a maximum change of 1 Hz to 5 Hz per time step, avoiding frequent and large-scale pump adjustments that could cause mechanical wear and flow fluctuations.

[0047] In a further embodiment, the system also includes a gas storage pre-pressurization module and a pulse ice-breaking module. Before the command receiving module receives the load reduction command from the power grid dispatch center, the gas storage pre-pressurization module sends a start signal to the air compressor connected to the gas pressure pipeline network, and raises the gas pressure in the gas storage tank connected to the air compressor to the upper limit of the safe pressure withstand threshold of the gas pressure pipeline network, then cuts off the power supply to the air compressor. This pre-pressurization is not about blindly maintaining high pressure 24 / 7, but rather about preferentially combining day-ahead electricity price forecasts, dispatch response probabilities, and park production plans to complete pressure storage before the period when demand response is highly likely to be triggered. The significance of this is that it schedules the operation of power-intensive air compressors during periods of parity or low electricity prices, pre-storing the aerodynamic potential energy that can be released upon restart, avoiding the need to temporarily start the air compressor during high-price or yield-reduction periods, thus offsetting transaction revenue.

[0048] The upper limit of the safe withstand pressure threshold is preferably set at 80% to 95% of the design withstand pressure value of the gas storage tank and gas pipeline network. Using this range ensures sufficient energy for subsequent pulse release while avoiding the adverse effects of prolonged top-level pressure accumulation on equipment lifespan and safety margins. Pressure sensors are preferably installed separately for the pressure inside the gas storage tank, the air compressor outlet pressure, and the pipeline network pressure, with cross-verification performed by the gas storage pre-pressurization module. If a pressure drop exceeding a set value, such as 5% to 15% of the upper limit of the safe withstand pressure threshold, is detected, and the pressure remains within the prepared response window, gas can be replenished once. This prevents insufficient gas storage during prolonged waiting periods for dispatch.

[0049] When the command execution module generates the power restoration command, the pulse ice-breaking module blocks the start-up control circuit of the main material pump connected to the target pipeline, and simultaneously sends an opening signal to the pneumatic solenoid valve connected in parallel with the target pipeline. This parallel connection can be understood as setting up a high-pressure gas injection branch connected to a gas storage tank upstream of the target pipeline, on its side wall, or at a dedicated pulse injection port. When the solenoid valve opens, high-pressure gas rushes into the target pipeline in a very short time, forming pressure pulses and oscillating waves. This oscillation effect does not rely on continuous purging to replace delivery, but rather uses short-duration, high-intensity impacts to disrupt the static frictional engagement between the crust and the pipe wall, inducing micro-cracks, localized debonding, and interface loosening within the crust. After this treatment, when the main pump restarts, it no longer faces a complete, firmly attached, continuously deposited solid shell, but rather a loosened, fractured, and weakened resistance layer, thus significantly reducing the starting peak current and mechanical load.

[0050] In one embodiment, the pulse icing module inputs the real-time crust thickness into the pneumatic pulse parameter matching template library. The pneumatic pulse parameter matching template library stores a mapping table between different crust thicknesses and their corresponding opening durations. The pulse icing module queries the template library to obtain the target opening duration. Upon receiving the opening signal, the pneumatic solenoid valve maintains its open state to release the high-pressure gas in the storage tank into the target pipeline, and closes after the target opening duration has elapsed. The template library uses a mapping table instead of a complex fluid impact model calculated in real-time on-site because industrial environments prioritize response speed, maintainability, and verifiability. The template library can be gradually established through multiple rounds of operational tests. During the tests, the icing effect, pressure wave attenuation, and subsequent main pump starting current peak are recorded for different crust thicknesses and opening durations. Parameters that balance effectiveness and safety are then selected as templates. Besides crust thickness, the template library can also be further grouped by pipe diameter, pipe material, and material category, but this does not change the core idea of ​​this invention: matching based on crust thickness as the primary key.

[0051] The target opening duration can be set from 0.05 seconds to 5 seconds, depending on the system volume and safety boundaries. For small to medium diameter pipes with thin crusts, 0.1 to 1 second is usually sufficient to generate a significant impact. For large diameter pipes or pipes with thick crusts, the duration can be appropriately extended, but not excessively, otherwise the pulse will evolve into a near-continuous gas injection, weakening the impact characteristics and potentially introducing unnecessary media disturbances. The peak pulse pressure is preferably limited to below the upper limit of the safe pressure withstand threshold of the gas pipeline network and controlled through the selection of flow-limiting orifice plates, quick-opening valves, and branch volume design. This ensures the pulse is controlled within a range that can disrupt crust adhesion without damaging the pipeline itself.

[0052] To ensure the main pump starts only after pulse ice breaking is complete, a time-delay relay is incorporated into the command execution module. The input of the time-delay relay receives the power restoration command generated by the command execution module, and its output connects to the start control terminal of the main material pump. The time-delay relay has an internal time-delay parameter, the value of which equals the target start duration plus a preset oscillation calming duration. After receiving the power restoration command, the time-delay relay waits for the time specified by the time-delay parameter before activating the start control circuit of the main material pump. The purpose of setting the oscillation calming duration is that after pulse gas injection, the pipeline experiences short-term pressure rebound and flow disturbance. If the main pump starts immediately before the oscillation has decayed, it may still cause superimposed impact. The preset oscillation calming duration can be set from 0.2 seconds to 10 seconds, with a suitable range of 0.5 seconds to 3 seconds. Specific values ​​can be calibrated based on pipeline length, medium compressibility, and pulse test results.

[0053] Regarding the overall system linkage, this invention can be executed in the following order. The park's energy and carbon intelligent control system pre-trains and deploys the phase change restart cost prediction model, the deep Q-network, and the reinforcement learning hovering control model. Before entering the high-probability demand response window, the gas storage pre-pressurization module uses periods of parity or low-price electricity to pressurize the gas storage tank to near the upper limit of the safe pressure withstand threshold, and causes the air compressor to exit the working state. When the power grid dispatch center issues a load reduction command, the command receiving module completes verification and issues an execution command. The status acquisition module cuts off the high-energy-consuming branch of the target pipeline, while keeping the sensor acquisition, communication control, and necessary micro-circulation branches online. Subsequently, the status acquisition module continuously collects dynamic viscosity, real-time center temperature, and ambient temperature; the phase change assessment module outputs the crust thickness and marginal restart cost in real time; the revenue prediction module outputs the marginal transaction revenue in real time; and the command execution module continuously compares the difference between the two. When the difference is still positive and continuously decreasing, the system continues to maintain shutdown and power supply; when the difference enters the zero threshold zone and meets the decreasing trend for three consecutive time steps, the system determines that the optimal recovery time has been reached. If the system is not equipped with a pulse ice-breaking module, it is preferable to restore the heating branch first, and then restore the main pump branch after the start-up conditions are met. If the system is equipped with a pulse ice-breaking module, the main pump start-up control circuit is kept open first, and then the pneumatic solenoid valve is triggered to release a high-voltage pulse according to the target opening duration given by the template library. After the time delay relay finishes timing, the main pump start-up control circuit is turned on, thereby achieving a smooth restart with near-zero surge or low surge.

[0054] In another embodiment, the micro-circulation frequency modulation module and the core benefit comparison control can operate simultaneously. That is, during the power outage and power-on period, the system does not passively wait for the crust to grow naturally, but instead uses a reinforcement learning hovering control model to continuously adjust the material circulation pump frequency, keeping the crust thickness as close as possible to the safe crust thickness threshold. This significantly slows down the growth rate of the marginal restart cost output by the phase change assessment module, and significantly extends the sustainable trading window corresponding to the benefit prediction module. The point at which the benefit and cost intersect is reached is later than in a completely static power outage scenario, thus securing more high-price electricity and carbon trading benefits for the industrial park. This is one of the key values ​​of this invention compared to simple shutdown control: instead of passively accepting physical constraints, it actively modulates the constraint boundary using weak flow and phase change mechanisms.

[0055] This invention can also incorporate multi-level protection logic to further enhance industrial feasibility. For example, when the real-time center temperature is below the process allowable lower limit, the crust thickness exceeds the safe crust thickness threshold by a certain percentage, the main pump's projected head is insufficient to cover the current resistance, the gas tank pressure is insufficient to support an effective pulse, or any critical sensor fails to connect, the system can directly skip the benefit comparison results and enter the recovery process or exit the power-on process ahead of time. The significance of this protection logic is that while this invention emphasizes optimal economics, safety boundaries always take precedence. For another example, when the phase change assessment module detects a sudden abnormal increase in the crust thickness growth rate, it can determine that the material composition or environmental conditions have deviated from the training distribution. In this case, a conservative mode can be automatically activated to shorten the comparison cycle and narrow the zero threshold band, reducing the risk of misjudgment.

[0056] In terms of implementation, this invention does not simply add a shutdown countdown to a traditional demand response system. Instead, it unifies the internal crusting process of pipelines, the increase in restart costs, and the marginal revenue process of the electricity carbon market into a single closed-loop control framework. Through a phase-change restart cost prediction model, the system can identify internal physical evolution, no longer treating industrial pipelines as ideal nodes that can be arbitrarily cut off and restored. Through a deep Q-network, the system can assess the future benefits of continued shutdown, no longer making static judgments based solely on a single instantaneous price. Through a difference comparator, the system can restore power at the critical point where revenue is just caught up with cost, maximizing the shutdown benefit zone. Through a micro-circulation frequency modulation module, the system further hovers the crust thickness near a safe margin, trading extremely low power consumption for a longer trading window. Through a gas storage pre-compression module, a pulse ice-breaking module, and a time-delay relay, the system transforms the most dangerous restart moment into a smooth process of ice breaking first, followed by pump startup, fundamentally reducing the risks of high current peaks and demand penalties.

[0057] The above embodiments are used to illustrate the technical concept of the present invention. For those skilled in the art, equivalent adjustments to the sensor type, number of network layers, sampling period, threshold range, pulse template details, and execution order without departing from the core idea of ​​the present invention can all fall within the protection scope of the present invention.

Claims

1. An energy and carbon intelligent control system for industrial parks, characterized in that, It includes an instruction receiving module, a status acquisition module, a phase change evaluation module, a profit prediction module, and an instruction execution module; The instruction receiving module receives load reduction instructions issued by the power grid dispatch center; Upon receiving the load reduction command, the status acquisition module cuts off the power supply to the target pipeline and collects real-time physical status data of the target material and ambient temperature data within the target pipeline through sensors. The phase change assessment module inputs the real-time physical state data and the ambient temperature data into the pre-trained phase change restart cost prediction model, and the pre-trained phase change restart cost prediction model outputs the real-time crust thickness and marginal restart cost of the inner wall of the target pipeline. The revenue prediction module acquires real-time electricity price data and real-time carbon trading price data, and inputs the real-time electricity price data and the real-time carbon trading price data into a pre-trained deep Q network. The pre-trained deep Q network outputs the marginal trading revenue for continuing to execute the load reduction command. The instruction execution module compares the marginal restart cost with the marginal transaction benefit. When the marginal restart cost equals the marginal transaction benefit, it generates a power restoration instruction and restores power to the target pipeline.

2. The energy and carbon intelligent control system for industrial parks as described in claim 1, characterized in that, The phase change restart cost prediction model includes a feature extraction network, a shell thickness prediction branch network, and a restart cost evaluation branch network. The feature extraction network receives the real-time physical state data and the ambient temperature data as input, and outputs a hidden layer feature vector. The crust thickness prediction branch network receives the hidden layer feature vector as input and outputs the real-time crust thickness of the inner wall of the target pipeline. The restart cost evaluation branch network receives the hidden layer feature vector and the real-time shell thickness as input, and outputs the marginal restart cost.

3. The energy and carbon intelligent control system for industrial parks as described in claim 2, characterized in that, The training process of the phase change restart cost prediction model includes: Acquire historical physical state data, historical ambient temperature data, historical actual crust thickness data, and historical actual restart power consumption data; The historical physical state data and the historical ambient temperature data are input into the phase change restart cost prediction model to be trained, and the predicted crust thickness and the predicted marginal restart cost are output. The first loss value is obtained by subtracting the predicted crust thickness from the historical actual crust thickness data. The second loss value is obtained by subtracting the predicted marginal restart cost from the historical actual restart power consumption data. The first loss value and the second loss value are summed to obtain the total loss value. The network parameters of the phase transition restart cost prediction model to be trained are updated using the backpropagation algorithm until the total loss value converges to below the convergence threshold, thus obtaining the trained phase transition restart cost prediction model.

4. The energy and carbon intelligent control system for industrial parks as described in claim 1, characterized in that, The real-time physical state data includes the dynamic viscosity data and real-time core temperature data of the target material. The marginal restart cost is the sum of the heat energy consumption for crust melting and the additional power consumption for pumping required to terminate the load reduction command and start the main material pump connected to the target pipeline at the current moment. The instruction execution module is equipped with a difference comparator. The difference comparator calculates the difference between the marginal transaction revenue and the marginal restart cost. When the difference shows a decreasing trend and equals zero within three consecutive time steps, the operation of generating the power restoration instruction is triggered.

5. The energy and carbon intelligent control system for industrial parks as described in claim 1, characterized in that, The energy and carbon intelligent control system for industrial parks also includes a micro-circulation frequency modulation module; The micro-circulation frequency modulation module maintains the power supply to the material circulation pump connected to the target pipeline during the power outage of the target pipeline; The micro-circulation frequency modulation module inputs the real-time shell thickness and the preset safe shell thickness threshold into the reinforcement learning hovering control model, and the reinforcement learning hovering control model outputs the target operating frequency. The micro-circulation frequency modulation module adjusts the operating frequency of the material circulation pump to the target operating frequency.

6. The energy and carbon intelligent control system for industrial parks as described in claim 5, characterized in that, The architecture of the reinforcement learning hovering control model adopts a deep deterministic policy gradient architecture, which includes an actor network and a critic network. The state space elements of the reinforcement learning hovering control model include the real-time crust thickness, the preset safe crust thickness threshold, and the operating frequency of the material circulation pump at the previous moment. The action space elements of the reinforcement learning hovering control model are the target running frequencies; The actor network outputs the target running frequency, and the critic network outputs the state-action value score corresponding to the target running frequency.

7. The energy and carbon intelligent control system for industrial parks as described in claim 6, characterized in that, The reinforcement learning hovering control model uses an environmental reward function to update parameters during the training phase. The environmental reward function includes a thickness maintenance reward term and an energy consumption penalty term; When the next moment is reached, if the absolute value of the difference between the real-time shell thickness obtained by the phase transition evaluation module and the preset safe shell thickness threshold is less than the preset difference threshold, the thickness maintenance reward item outputs a positive reward. The value of the energy consumption penalty term is inversely proportional to the value of the target operating frequency; The environmental reward function sums the thickness maintenance reward term and the energy consumption penalty term to obtain a total reward value, and uses the total reward value to update the weight parameters of the actor network and the critic network.

8. The energy and carbon intelligent control system for industrial parks as described in claim 1, characterized in that, The energy and carbon intelligent control system for industrial parks also includes a gas storage pre-compression module and a pulse ice-breaking module. During the time period before the command receiving module receives the load reduction command issued by the power grid dispatch center, the gas storage pre-compression module sends a start signal to the air compressor connected to the gas pressure pipeline network, and raises the gas pressure of the gas storage tank connected to the air compressor to the upper limit of the safe pressure withstand threshold of the gas pressure pipeline network, and then cuts off the power supply to the air compressor. The pulse ice-breaking module blocks the starting power circuit of the main material pump connected to the target pipeline when the instruction execution module generates the power restoration instruction, and at the same time sends an opening signal to the pneumatic solenoid valve connected in parallel with the target pipeline.

9. The energy and carbon intelligent control system for industrial parks as described in claim 8, characterized in that, The pulse ice-breaking module inputs the real-time crust thickness into the pneumatic pulse parameter matching template library; The pneumatic pulse parameter matching template library stores a mapping table between different shell thicknesses and corresponding opening durations. The pulse ice-breaking module retrieves the target activation duration from the pneumatic pulse parameter matching template library; After receiving an opening signal, the pneumatic solenoid valve remains open to release the high-pressure gas in the gas storage tank into the target pipeline, and closes after the target opening duration has elapsed.

10. The energy and carbon intelligent control system for industrial parks as described in claim 9, characterized in that, The instruction execution module is equipped with a time-delay relay. The input terminal of the time delay relay receives the power restoration command generated by the command execution module, and the output terminal of the time delay relay is connected to the start control terminal of the main material pump; The time delay relay has a time delay parameter inside, and the value of the time delay parameter is equal to the target activation duration plus a preset oscillation calming duration; After receiving the power restoration command for the specified time period, the time delay relay activates the starting power circuit of the main material pump.

Images