An intelligent scheduling method for cross-seasonal thermal storage based on thermodynamic physical mechanism constraints

Abstract

This invention discloses a cross-seasonal intelligent scheduling method for thermal energy storage based on thermodynamic physical mechanism constraints, comprising the following steps: constructing a thermodynamically aware cross-spatial feature extraction network to output explicit thermodynamic state vectors; integrating a causal world model constrained by the first and second laws of thermodynamics, using the second law of thermodynamics to forcibly cut off false causal paths in the entropy reduction direction, achieving physically self-consistent counterfactual reasoning; an adaptive decision takeover engine based on operating condition awareness, pursuing economy under normal operating conditions and forcibly switching to a symbolically constrained safety mode under extreme operating conditions; and hierarchical federated reinforcement learning based on thermodynamic constraint awareness to achieve cross-site knowledge transfer. This invention eliminates the state prediction divergence problem in long-cycle iterations, achieves physical self-consistency throughout the scheduling process, and eliminates the safety risks of equipment operating beyond limits.

Description

Technical Field

[0001] This invention relates to the field of intelligent scheduling and control technology, specifically to a cross-seasonal intelligent scheduling method for thermal energy storage based on thermodynamic physical mechanism constraints. Background Technology

[0002] Under the "dual-carbon" strategy, the installed capacity of renewable energy sources such as wind power and photovoltaics continues to grow. However, their intermittent and fluctuating characteristics pose significant challenges to the safe and stable operation of the power grid. Meanwhile, there is a significant seasonal supply-demand mismatch in centralized heating in northern cities, with surplus renewable energy in summer and a large heating energy shortage in winter. Cross-seasonal thermal storage technology can convert surplus renewable energy in summer into thermal energy for storage and release during the winter heating season, making it one of the core technologies for solving the problem of renewable energy consumption and seasonal heating supply-demand mismatch. Among these technologies, high-temperature solid particle (ceramic particles, quartz sand, blast furnace slag, etc.) thermal storage technology has advantages such as high storage temperature (600℃-1000℃), high energy density, low cost, long lifespan, and environmental friendliness, making it the mainstream technology for large-capacity cross-seasonal thermal storage.

[0003] Cross-seasonal thermal energy storage scheduling differs fundamentally from conventional short-term energy storage and district heating system scheduling, and faces the following technical challenges: 1. During the charging and releasing process of high-temperature solid particles at 600℃-1000℃ in a gravity-flow thermal storage bed, there is intense gas-solid phase-to-phase heat transfer and nonlinear thermal radiation effects. Their thermodynamic characteristics exhibit strong nonlinearity, which cannot be accurately described by traditional linear models. In the long-term cumulative iteration across seasons, catastrophic state prediction divergence will occur, causing the scheduling strategy to completely fail.

[0004] 2. Cross-seasonal thermal storage requires planning thermal storage capacity on a long-term scale of several months, while responding to grid frequency regulation needs on a second-minute scale. Thermodynamic state transitions have strong path dependence, and traditional scheduling methods cannot meet the dual requirements of long-term energy balance and short-term rapid response.

[0005] 3. Extreme weather events such as cold waves and warm winters can cause sudden changes in heat load exceeding 30%. Traditional scheduling methods rely on historical data fitting and lack counterfactual reasoning capabilities. They are unable to cope with extreme conditions that have not occurred in the past and are prone to the risk of premature depletion of heat storage capacity and heat supply interruption.

[0006] 4. Multiple uncertainties such as weather, electricity price, power grid frequency, and heat load are coupled together, and the stationarity assumption of the traditional Markov decision process is no longer applicable, resulting in a situation where the robustness and economy of the scheduling strategy cannot be taken into account at the same time. Summary of the Invention

[0007] The present invention proposes a cross-seasonal intelligent scheduling method for thermal energy storage based on thermodynamic physical mechanism constraints, which can at least solve one of the technical problems in the background art.

[0008] To achieve the above objectives, the present invention adopts the following technical solution: A cross-seasonal intelligent thermal energy storage scheduling method based on thermodynamic physical mechanism constraints, targeting the cross-seasonal thermal energy storage scheduling scenario of a high-temperature solid particle gravity flow thermal energy storage bed at 600℃-1000℃, includes the following steps: S1: Simultaneously collects multi-source time-series data from meteorology, electricity prices, power grids, and heating networks. After thermodynamic normalization preprocessing, it outputs an explicit four-dimensional thermodynamic state vector. ,in The average temperature of the thermal storage. In a charged state, For the rate of temperature change, The rate of change of state of charge; S2: Feature extraction and long-term situation characterization of preprocessed multimodal data are performed through a three-level cascaded cross-temporal feature extraction network. The cross-temporal feature extraction network includes an intraday dynamic response extractor, a short-term weather process extractor, and a cross-seasonal long-term feature extractor. The extractors at each level are physically consistent through a thermodynamic state checker to ensure that the output features can be mapped to the real physical state without bias. S3: Based on the physical mechanism-data dual-driven topology recognition engine, a causal directed acyclic graph of the thermal storage system is constructed. By using a causal mask based on the second law of thermodynamics, false causal paths in the entropy reduction direction are forcibly cut off. At the same time, the thermal balance equation of the first law of thermodynamics is embedded into the state transition function, so that the trajectory of the model state deduction is anchored on the manifold of the thermodynamic physical equation, forming a physically self-consistent causal world model. S4: Based on the causal world model, perform interventionist counterfactual reasoning to predict the thermodynamic state evolution trajectory and constraint satisfaction of different scheduling strategies under extreme weather and load change conditions, and output the state deviation and uncertainty quantification results of different strategies; S5: Based on the pre-trained large language model of injected thermal storage scheduling knowledge, candidate scheduling strategies are generated by combining the current thermodynamic state and counterfactual reasoning results. During the generation process, the candidate strategies are initially physically verified by using a thermodynamic symbol knowledge base. S6: The dynamic fusion weight of the neural solver and the symbolic verifier is calculated by the condition-aware regulator. Under normal operating conditions, the neural solver is used first to quickly optimize the scheduling strategy. Under extreme approximation conditions, the system is forced to switch to the symbolic verifier to solve the hard constraints and output the optimal scheduling action that satisfies the thermodynamic and physical constraints. S7: Calculates the thermodynamic state prediction variance of the scheduling strategy based on Monte Carlo sampling results using counterfactual reasoning. When the variance exceeds a preset threshold, a conservative scheduling strategy is automatically triggered, tightening the storage temperature. Safety constraints with power ramp-up.

[0009] As a preferred embodiment of the cross-seasonal thermal energy storage intelligent scheduling method based on thermodynamic physical mechanism constraints described in this invention, wherein: in step S2, the time window of the intraday dynamic response extractor is 24 hours, used to capture minute-level meteorological fluctuations and the thermal power throughput capacity of the thermal energy storage system under grid frequency regulation commands; the time window of the short-cycle weather process extractor is 7 days, used to identify the cumulative impact of continuous weather processes on heat load and thermal energy storage status; the time window of the cross-seasonal long-cycle feature extractor is 30 days or more, used to track the drift of basic heat load and the gradual change of ambient temperature caused by seasonal alternation; the thermodynamic state verifier includes a temperature derivative comparison unit, a SOC change rate comparison unit, and an anomaly feature correction unit, when the extracted temperature change rate deviates from the theoretical value of the heat balance equation by more than 0.5℃ / h, or When the rate of change deviates from the theoretical value of the law of conservation of thermal storage capacity by more than 0.1% / min, anomaly characteristic correction is triggered. The correction formula is as follows:

[0010] in, The preset confidence weights.

[0011] As a preferred embodiment of the cross-seasonal thermal energy storage intelligent scheduling method based on thermodynamic physical mechanism constraints described in this invention, wherein: in step S3, the node set of the causal directed acyclic graph ,in For meteorological conditions, For electricity price signals, For power grid status, For thermal storage SOC, For heat load, For scheduling actions, The average temperature of the thermal storage; the causal edge set includes , , , , , , The causal mask forcibly preserves core causal edges that follow the first law of thermodynamics, forcibly severs false causal edges that violate the second law of thermodynamics in the direction of entropy reduction, prohibits causal paths where low-temperature heat sources spontaneously transfer heat to high-temperature energy storage bodies, and simultaneously prohibits reverse causal paths. , The discretized expression of the state transition function is:

[0012] In the formula: The temperature at the next time step after discretization; Let be the discrete time step; The total heat capacity of the heat storage medium; For heat charging efficiency; This refers to the heating power; The overall heat transfer coefficient of the system; Ambient temperature; This represents the heat release power.

[0013] As a preferred embodiment of the cross-seasonal thermal energy storage intelligent scheduling method based on thermodynamic physical mechanism constraints described in this invention, wherein: in step S4, the counterfactual reasoning is performed through interference calculations... The implementation and inference expression are as follows:

[0014] In the formula: It is a counterfactual reasoning operator; The thermodynamic state vector, This represents the current system state. As the baseline action, This is an intervention action.

[0015] As a preferred embodiment of the cross-seasonal thermal energy storage intelligent scheduling method based on thermodynamic physical mechanism constraints described in this invention, wherein: in step S5, the thermal energy storage scheduling domain knowledge is implemented through a three-level injection method: First, a domain knowledge base is constructed, which includes thermodynamic physical equations, expert experience rules, and historical scheduling cases. ; Secondly, the domain knowledge base is injected into the pre-trained large language model through system prompts to clarify the physical constraints of the scheduling task. Finally, the large language model is efficiently fine-tuned using a low-rank adapter (LoRA), freezing the weights of the basic model and updating only the domain adaptation parameters of the attention layer bypass. The LoRA has a rank of 16 and an alpha of 32. The thermodynamic feasibility check is performed by analyzing the control parameters of the candidate strategy, calculating the state transition trajectory based on the thermal balance equation, and verifying whether the temperature and SOC are within a safe range. If the check fails, a conservative backoff strategy is triggered.

[0016] In the formula: Safety action strategy for reversal; For a predefined safe action space; The original candidate actions generated for the large language model.

[0017] As a preferred embodiment of the cross-seasonal thermal energy storage intelligent scheduling method based on thermodynamic physical mechanism constraints described in this invention, wherein: in step S6, the normal operating condition is defined as the state safety degree. And weighted violation frequency The limit approximation condition is defined as follows: or ;in The calculation formula is based on the degree to which the current physical state of the system approaches the thermodynamic safety dead zone.

[0018] in: These are the preset physical limits for temperature and state of charge, respectively. The violation frequency statistics unit calculates the weighted violation frequency. ; Element regulator according to and Dynamic output fusion weights Among them, when and The system is currently in normal operating condition. Neural solvers are preferred; when or The system determines that it is in a limit approximation condition. The symbolic verifier is forced to be used, and the hybrid optimization objective function is:

[0019] In the formula: For the final optimized scheduling action; This is the loss function for the neural module; For symbolic knowledge base The constraint loss function; These are thermodynamic prior parameters.

[0020] As a preferred embodiment of the cross-seasonal thermal energy storage intelligent scheduling method based on thermodynamic physical mechanism constraints described in this invention, wherein: the thermodynamic symbol knowledge base It includes three types of core symbolic equations: The first category consists of symbols for the heat balance equation. The expression is:

[0021] Used to constrain the energy conservation of thermal storage systems; The second category is SOC-defined equation symbols. The expression is:

[0022] in, This represents the system's maximum thermal storage capacity, used to map temperature to the percentage of thermal storage capacity. The third category consists of symbols for safety and economic dispatch constraint equations. With the goal of minimizing time-of-use electricity costs, the constraints include:

[0023]

[0024]

[0025] in, This is the maximum power adjustment step size allowed by the system.

[0026] As a preferred embodiment of the cross-seasonal thermal energy storage intelligent scheduling method based on thermodynamic physical mechanism constraints described in this invention, the method further includes a thermodynamic constraint-aware federated hierarchical collaborative scheduling step, employing a three-layer federated architecture of global coordination layer - regional coordination layer - local execution layer; the local execution layer uses a proximal policy optimization algorithm with thermodynamic constraint penalty terms to iterate local policies and prune parameters. Generalized dominance estimation coefficient When the strategy deduction triggers a thermodynamic limit breach, a circuit breaker mechanism is triggered to cut off reward feedback; when the regional coordination layer and the global coordination layer aggregate model parameters, the thermodynamic constraint satisfaction of each local node is used as the core aggregation weight, and the aggregation formula is:

[0027] In the formula: These are global model parameters; For the region Model parameters; For regional aggregation weights; in, , For the region The number of samples, This represents the degree of satisfaction of thermodynamic constraints.

[0028] As a preferred embodiment of the cross-seasonal intelligent scheduling method for thermal storage based on thermodynamic physical mechanisms described in this invention, the method further includes thermodynamic safety monitoring and AI decision-making fault-tolerant steps. This involves real-time acquisition of five core safety parameters of the thermal storage system: average temperature, state of charge (SOC), tank pressure, temperature change rate, and SOC change rate. Graded safety thresholds are set; when a parameter is within a warning range, power limiting is triggered; when a parameter is within a critical range, the current scheduling command is immediately stopped, a safety protection program is initiated, and the system switches to manual control mode. A hardware-level hard constraint protection unit, independent of the AI ​​decision chain, is set up; when the thermal storage temperature exceeds the hardware safety cut-off threshold... SOC is lower than Or pressure exceeds In some cases, an emergency shutdown action is triggered directly without going through the AI ​​decision-making chain.

[0029] As a preferred embodiment of the cross-seasonal thermal energy storage intelligent scheduling method based on thermodynamic physical mechanism constraints described in this invention, wherein: in step S7, the constraint boundary of the conservative scheduling strategy is: The safety lower limit is set as follows: The lower limit of safe temperature is set to The upper limit of temperature safety is set to The maximum power adjustment step size is tightened to... Conservative scheduling strategy to maximize the next time step The expected value is the optimization objective, expressed as: The constraints are .

[0030] The beneficial effects of this invention are: This invention embeds thermodynamic physical laws into the entire process of feature extraction and state modeling. A three-level spatiotemporal feature extraction network combined with a thermodynamically constrained causal world model eliminates the state prediction divergence problem in long-period iterations. This invention achieves physical self-consistency throughout the scheduling process through a triple protection mechanism of full-process thermodynamic constraint verification, neural symbol hybrid optimization hard constraint interception, and hardware-level security protection, thus eliminating the safety risks of equipment operating beyond limits. Attached Figure Description

[0031] Figure 1 This is a flowchart illustrating the overall steps of the intelligent scheduling method for cross-seasonal thermal energy storage based on thermodynamic physical mechanisms, as described in this invention.

[0032] Figure 2 This is a panoramic technical architecture diagram of the cross-seasonal thermal energy storage intelligent scheduling method based on thermodynamic physical mechanism constraints of the present invention. Detailed Implementation

[0033] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are some embodiments of the present invention, but not all embodiments.

[0034] like Figures 1-2 As shown, a cross-seasonal intelligent thermal energy storage scheduling method based on thermodynamic physical mechanism constraints is provided, targeting the cross-seasonal thermal energy storage scheduling scenario of a high-temperature solid particle gravity flow thermal energy storage bed of 600℃-1000℃, including the following steps: S1: Multimodal situational awareness and thermodynamic state coding steps: Simultaneously collect multi-source time-series data from meteorology, electricity prices, power grids, and heating networks. After thermodynamic normalization preprocessing, output an explicit four-dimensional thermodynamic state vector. ,in The average temperature of the thermal storage. In a charged state, For the rate of temperature change, The rate of change of state of charge; Specifically, the multi-source time-series data includes: meteorological data (irradiance, ambient temperature, wind speed, precipitation probability), electricity price data (time-of-use pricing, real-time spot pricing), power grid data (frequency deviation, line power flow, power curtailment indication), and heating network data (supply water temperature, return water temperature, flow rate, thermal storage SOC, thermal power); the thermodynamic normalization preprocessing formula is:

[0035] In the formula: This is the normalized thermodynamic input vector; This is the original input data; For reference working point; Within the physically feasible range; This is a validity indicator vector; S2: The preprocessed multimodal data is subjected to feature extraction and long-term situation characterization through a three-level cascaded cross-temporal feature extraction network. The cross-temporal feature extraction network includes an intraday dynamic response extractor, a short-term weather process extractor, and a cross-seasonal long-term feature extractor. The extractors at each level are physically consistent through a thermodynamic state checker to ensure that the output features can be mapped to the real physical state without bias. Specifically, the core configuration and functions of the three-stage extractor are as follows: Intraday Dynamic Response Extractor: Time window L=24h, encoder layer 2 layers, attention head 8 layers, hidden layer dimension 512, used to capture the thermal power throughput capacity limit and response delay of thermal storage system under minute-level meteorological fluctuations, power grid frequency regulation commands, and real-time electricity price signals, and output intraday charging and discharging capacity adequacy characteristics. Short-cycle weather process extractor: time window L=7d, encoder layer 2 layers, attention head 8 layers, hidden layer dimension 512, used to identify the cumulative impact of weather processes (continuous rain, cold wave) on heat load and heat storage status within a 7-day rolling window, and to establish a cross-day heat accumulation and dissipation model. Long-term cross-seasonal feature extractor: Time window L=30 days and above, 2 encoder layers, 8 attention heads, 512 hidden layer dimensions, used to track the drift of basic heat load and gradual change of ambient temperature caused by seasonal changes, and extract the long-term natural heat dissipation pattern of the system and the energy transfer demand across months.

[0036] The thermodynamic state verifier includes a temperature derivative comparison unit, a SOC change rate comparison unit, a verification result buffering unit, and an anomaly feature correction unit. Temperature derivative comparison unit: compares the deviation between the extracted rate of temperature change and the theoretical value of the heat balance equation. The calculation formula is as follows:

[0037] When the deviation exceeds the first threshold A verification failure signal is triggered at the time; SOC Change Rate Comparison Unit: This unit compares the deviation between the extracted SOC change rate and the theoretical value based on the law of conservation of thermal storage capacity. The calculation formula is as follows:

[0038] When the deviation exceeds the second threshold A verification failure signal is triggered at the time; Anomaly correction unit: When verification fails, anomalies are corrected using a weighted average method. The correction formula is as follows:

[0039] in, The preset confidence weights.

[0040] S3: Based on the physical mechanism-data dual-driven topology recognition engine, a causal directed acyclic graph of the thermal storage system is constructed. By using a causal mask based on the second law of thermodynamics, false causal paths in the entropy reduction direction are forcibly cut off. At the same time, the thermal balance equation of the first law of thermodynamics is embedded into the state transition function, so that the trajectory of the model state deduction is anchored on the manifold of the thermodynamic physical equation, forming a physically self-consistent causal world model. Specifically, the node set of a causal directed acyclic graph ,in For meteorological conditions, For electricity price signals, For power grid status, For thermal storage SOC, For heat load, For scheduling actions, This represents the average temperature of the thermal storage.

[0041] The set of causal edges includes , , , , , , Causal masking forcibly preserves core causal edges that follow the first law of thermodynamics, forcibly severs false causal edges that violate the second law of thermodynamics in the direction of entropy reduction, prohibits causal paths where low-temperature heat sources spontaneously transfer heat to high-temperature energy storage bodies, and simultaneously prohibits reverse causal paths. , The discretized expression for the state transition function is:

[0042] In the formula: The temperature at the next time step after discretization; Let be the discrete time step; The total heat capacity of the heat storage medium; For heat charging efficiency; This refers to the heating power; The overall heat transfer coefficient of the system; Ambient temperature; This refers to the heat release power; Reward function (combining economics and security)

[0043] In the formula: For instant rewards; The current electricity price; This represents the change in the thermal storage state. For reward weighting.

[0044] Observation function

[0045] In the formula: These are system observations; and These are the linear parameters of the observation model.

[0046] S4: Based on the causal world model, perform intervention-based counterfactual reasoning to predict the thermodynamic state evolution trajectory and constraint satisfaction of different scheduling strategies under extreme weather and load change conditions, and output the state deviation and uncertainty quantification results of different strategies; Specifically, counterfactual reasoning involves interfering with budgets. The implementation and inference expression are as follows:

[0047] In the formula: It is a counterfactual reasoning operator; The thermodynamic state vector, This represents the current system state. As the baseline action, This is an intervention action.

[0048] The uncertainty quantification method includes the following steps: Step 1: Perform Monte Carlo sampling based on the causal world model, number of sampling times. :

[0049] In the formula: For the first The predicted state of the next sample; This represents the total number of samples taken.

[0050] Step 2: Calculate the variance of the state prediction:

[0051] In the formula: The variance of the state prediction; This is the mean of the predicted state.

[0052] in This is the mean of the state prediction.

[0053] Step 3: Calculate the uncertainties in temperature and SOC:

[0054] In the formula: These are the standard deviations of temperature and state of charge, respectively.

[0055] S5: Based on the pre-trained large language model of injected thermal storage scheduling knowledge, candidate scheduling strategies are generated by combining the current thermodynamic state and counterfactual reasoning results. During the generation process, the candidate strategies are initially physically verified by using a thermodynamic symbol knowledge base. Specifically, knowledge in the field of thermal energy storage scheduling is implemented through a three-level injection method: First, a domain knowledge base is constructed, which includes thermodynamic physical equations, expert experience rules, and historical scheduling cases. ; Secondly, the domain knowledge base is injected into the pre-trained large language model through system prompts to clarify the physical constraints of the scheduling task. Finally, the large language model is efficiently fine-tuned using a low-rank adapter (LoRA), freezing the weights of the base model and updating only the domain adaptation parameters of the attention layer bypass. The LoRA has a rank of 16 and an alpha of 32. Thermodynamic feasibility is verified by analyzing the control parameters of the candidate strategy, calculating the state transition trajectory based on the heat balance equation, and verifying whether the temperature and SOC are within a safe range. If the verification fails, a conservative backoff strategy is triggered.

[0056] In the formula: Safety action strategy for reversal; For a predefined safe action space; The original candidate actions generated for the large language model.

[0057] S6: The dynamic fusion weight of the neural solver and the symbolic verifier is calculated by the condition-aware regulator. Under normal operating conditions, the neural solver is used first to quickly optimize the scheduling strategy. Under extreme approximation conditions, the system is forced to switch to the symbolic verifier to solve the hard constraints and output the optimal scheduling action that satisfies the thermodynamic and physical constraints. Specifically, normal operating conditions are defined as state safety. And weighted violation frequency The limit approximation condition is defined as follows: or ;in The calculation formula is based on the degree to which the current physical state of the system approaches the thermodynamic safety dead zone.

[0058] in: These are the preset physical limits for temperature and state of charge, respectively. The violation frequency statistics unit calculates the weighted violation frequency. ; Element regulator according to and Dynamic output fusion weights Among them, when and The system is currently in normal operating condition. Neural solvers are preferred; when or The system determines that it is in a limit approximation condition. The symbolic verifier is forced to be used, and the hybrid optimization objective function is:

[0059] In the formula: For the final optimized scheduling action; This is the loss function for the neural module; For symbolic knowledge base The constraint loss function; These are thermodynamic prior parameters.

[0060] Thermodynamics Symbol Knowledge Base It includes three types of core symbolic equations: The first category consists of symbols for the heat balance equation. The expression is:

[0061] Used to constrain the energy conservation of thermal storage systems; The second category is SOC-defined equation symbols. The expression is:

[0062] in, This represents the system's maximum thermal storage capacity, used to map temperature to the percentage of thermal storage capacity. The third category consists of symbols for safety and economic dispatch constraint equations. With the goal of minimizing time-of-use electricity costs, the constraints include:

[0063]

[0064]

[0065] in, This is the maximum power adjustment step size allowed by the system.

[0066] S7: Calculates the thermodynamic state prediction variance of the scheduling strategy based on Monte Carlo sampling results using counterfactual reasoning. When the variance exceeds a preset threshold, a conservative scheduling strategy is automatically triggered, tightening the storage temperature. Safety constraints related to power ramp-up; Specifically, the conservative scheduling strategy triggering mechanism: when or or When this happens, a conservative scheduling strategy is automatically triggered:

[0067] In the formula: A conservative scheduling strategy; This represents the expected state of charge at the next moment.

[0068] Characteristics of a conservative strategy: 1. Prioritize keeping the SOC within safe limits ( ) 2. Reduce power change rate limit ( kW / min 3. Increase safety margin ( C) 4. Avoid aggressive scheduling under extreme operating conditions.

[0069] This invention underwent a 12-month field test at three thermal storage heating power stations in North China. The verification data are as follows: The experimental site configuration is shown in Table 1 below: Table 1

[0070] The quantitative verification data are shown in Tables 2 and 3 below: Table 2: Overall Performance Comparison over 12 Months

[0071] Table 3: Extreme Weather Scenario Tests

[0072] Statistical analysis results: Using the Wilcoxon signed-rank test, the present invention shows statistical significance in scheduling cost (p < 0.001) compared to traditional MPC; and the improvement in thermal constraint violation rate (p < 0.0001).

[0073] Specific application examples are as follows: Example 1: Thermal energy storage scheduling during a cold wave: Scenario description: A strong cold wave will hit in the next 72 hours, with the ambient temperature expected to drop from 0°C to 15°C, lasting for about 48 hours.

[0074] Data input: Weather forecast: Temperature curve for the next 72 hours; Electricity price forecast: Time-of-use electricity price curve; Initial state of thermal storage: C, Heat load forecast: Heat load curve for the next 72 hours; Long-term situational simulation processing: 1. Multimodal input: meteorological data (irradiance, temperature, wind speed), electricity price data, thermal storage status, heat load; 2. Intraday encoder: processes a 24-hour window to capture intraday temperature fluctuations; 3. Short-cycle encoder: processes L=7d windows to capture the evolution trend of cold waves; 4. Thermodynamic normalization: Normalize the temperature input to a thermodynamically feasible range; Causal world model predictions: 1. State transition prediction: Predicting changes in thermal storage temperature and SOC based on the heat balance equation; 2. Extreme weather impact assessment: Predicting the additional heat load caused by cold waves; 3. Risk assessment: Calculate thermodynamic uncertainties; Strategy generation and optimization: 1. Generate candidate strategies based on large-scale generative language models; 2. Thermodynamic constraint verification: Check whether the SOC and temperature meet the constraints; 3. Neural symbol mixing optimization: Selecting the optimal strategy; Optimal scheduling scheme: Phase 1 (24 hours before the cold wave): During the off-peak electricity price period from 22:00 to 06:00, charge at 400kW to increase the SOC to 85%; Phase 2 (during cold wave): Maintain SOC above 80%, minimize charging and discharging heat, and only respond to frequency regulation needs; Phase 3 (after the cold wave ends): Gradually restore SOC to 75% during periods of low electricity prices; Example 2: Multi-region federated collaborative scheduling: Scenario Description: Three thermal energy storage power plants in three regions are subject to federal collaborative optimization. Each region has different load characteristics and electricity pricing policies.

[0075] Regional configuration: Area A: Urban heating area, with stable load and small peak-valley electricity price difference; Area B: Industrial user area, with large load fluctuations, participating in the spot market; Region C: High penetration area of ​​new energy sources, with a relatively high rate of wind and solar curtailment; Federated learning process: 1. The local near-end policy optimization algorithm for each region is updated, and the reward function includes thermodynamic constraints; 2. Regional Coordination Layer Convergence: Ensuring thermodynamic consistency within the region; 3. Global Coordination Layer Aggregation: Transferring Causal Knowledge Structure; 4. Thermodynamic consistency verification: Check whether the post-polymerization strategy meets thermodynamic constraints; Synergistic effect: The stable load characteristics of region A help region B cope with sudden load fluctuations; The surplus thermal storage capacity of region C is used to absorb the curtailment of wind and solar power in region B. The global thermodynamic constraint satisfaction rate increased from 87% to 96%.

[0076] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0077] The various embodiments in this specification are described in a related manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the system embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0078] 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 do 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. A method for intelligent scheduling of cross-seasonal thermal storage based on thermodynamic physical mechanism constraints, for the cross-seasonal thermal storage scheduling scenario of a high-temperature solid particle gravity flow thermal storage bed with a temperature of 600-1000℃, characterized in that, Includes the following steps: S1: Simultaneously collects multi-source time-series data from meteorology, electricity prices, power grids, and heating networks. After thermodynamic normalization preprocessing, it outputs an explicit four-dimensional thermodynamic state vector. ,in The average temperature of the thermal storage. In a charged state, For the rate of temperature change, The rate of change of state of charge; S2: Feature extraction and long-term situation characterization of preprocessed multimodal data are performed through a three-level cascaded cross-temporal feature extraction network. The cross-temporal feature extraction network includes an intraday dynamic response extractor, a short-term weather process extractor, and a cross-seasonal long-term feature extractor. The extractors at each level are physically consistent through a thermodynamic state checker to ensure that the output features can be mapped to the real physical state without bias. S3: Based on the physical mechanism-data dual-driven topology recognition engine, a causal directed acyclic graph of the thermal storage system is constructed. By using a causal mask based on the second law of thermodynamics, false causal paths in the entropy reduction direction are forcibly cut off. At the same time, the thermal balance equation of the first law of thermodynamics is embedded into the state transition function, so that the trajectory of the model state deduction is anchored on the manifold of the thermodynamic physical equation, forming a physically self-consistent causal world model. S4: Based on the causal world model, perform interventionist counterfactual reasoning to predict the thermodynamic state evolution trajectory and constraint satisfaction of different scheduling strategies under extreme weather and load change conditions, and output the state deviation and uncertainty quantification results of different strategies; S5: Based on the pre-trained large language model of injected thermal storage scheduling knowledge, candidate scheduling strategies are generated by combining the current thermodynamic state and counterfactual reasoning results. During the generation process, the candidate strategies are initially physically verified by using a thermodynamic symbol knowledge base. S6: The dynamic fusion weight of the neural solver and the symbolic verifier is calculated by the condition-aware regulator. Under normal operating conditions, the neural solver is used first to quickly optimize the scheduling strategy. Under extreme approximation conditions, the system is forced to switch to the symbolic verifier to solve the hard constraints and output the optimal scheduling action that satisfies the thermodynamic and physical constraints. S7: Calculates the thermodynamic state prediction variance of the scheduling strategy based on Monte Carlo sampling results using counterfactual reasoning. When the variance exceeds a preset threshold, a conservative scheduling strategy is automatically triggered, tightening the storage temperature. Safety constraints with power ramp-up.

2. The cross-seasonal thermal energy storage intelligent scheduling method based on thermodynamic physical mechanism constraints according to claim 1, characterized in that: In step S2, the time window of the intraday dynamic response extractor is 24 hours, used to capture minute-level meteorological fluctuations and the thermal power throughput capacity of the thermal storage system under grid frequency regulation commands; the time window of the short-cycle weather process extractor is 7 days, used to identify the cumulative impact of continuous weather processes on heat load and thermal storage status; the time window of the cross-seasonal long-cycle feature extractor is 30 days or more, used to track the drift of basic heat load and the gradual change of ambient temperature caused by seasonal alternation; the thermodynamic state verifier includes a temperature derivative comparison unit, a SOC change rate comparison unit, and an anomaly feature correction unit. When the extracted temperature change rate deviates from the theoretical value of the heat balance equation by more than 0.5℃ / h, or When the rate of change deviates from the theoretical value of the law of conservation of thermal storage capacity by more than 0.1% / min, anomaly characteristic correction is triggered. The correction formula is as follows: in, The preset confidence weights.

3. The cross-seasonal intelligent thermal energy storage scheduling method based on thermodynamic physical mechanism constraints according to claim 1, characterized in that: In step S3, the node set of the causal directed acyclic graph ,in For meteorological conditions, For electricity price signals, For power grid status, For thermal storage SOC, For heat load, For scheduling actions, The average temperature of the thermal storage; the causal edge set includes , , , , , , The causal mask forcibly preserves core causal edges that follow the first law of thermodynamics, forcibly severs false causal edges that violate the second law of thermodynamics in the direction of entropy reduction, prohibits causal paths where low-temperature heat sources spontaneously transfer heat to high-temperature energy storage bodies, and simultaneously prohibits reverse causal paths. , The discretized expression of the state transition function is: In the formula: The temperature at the next time step after discretization; Let be the discrete time step; The total heat capacity of the heat storage medium; For heat charging efficiency; This refers to the heating power; The overall heat transfer coefficient of the system; Ambient temperature; This represents the heat release power.

4. The cross-seasonal intelligent scheduling method for thermal energy storage based on thermodynamic physical mechanism constraints according to claim 1, characterized in that: In step S4, the counterfactual reasoning is achieved through interference with the budget. The implementation and inference expression are as follows: In the formula: It is a counterfactual reasoning operator; The thermodynamic state vector, This represents the current system state. As the baseline action, This is an intervention action.

5. The cross-seasonal intelligent thermal energy storage scheduling method based on thermodynamic physical mechanism constraints according to claim 1, characterized in that: In step S5, the knowledge of thermal energy storage scheduling is implemented through a three-level injection method: First, a domain knowledge base is constructed, which includes thermodynamic physical equations, expert experience rules, and historical scheduling cases. ; Secondly, the domain knowledge base is injected into the pre-trained large language model through system prompts to clarify the physical constraints of the scheduling task. Finally, the large language model is efficiently fine-tuned using a low-rank adapter (LoRA), freezing the weights of the basic model and updating only the domain adaptation parameters of the attention layer bypass. The LoRA has a rank of 16 and an alpha of 32. The thermodynamic feasibility check is performed by analyzing the control parameters of the candidate strategy, calculating the state transition trajectory based on the thermal balance equation, and verifying whether the temperature and SOC are within a safe range. If the check fails, a conservative backoff strategy is triggered. In the formula: Safety action strategy for reversal; For a predefined safe action space; The original candidate actions generated for the large language model.

6. The cross-seasonal intelligent scheduling method for thermal energy storage based on thermodynamic physical mechanism constraints according to claim 1, characterized in that: In step S6, the normal operating condition is defined as the state safety level. And weighted violation frequency The limit approximation condition is defined as follows: or ;in The calculation formula is based on the degree to which the current physical state of the system approaches the thermodynamic safety dead zone. in: These are the preset physical limits for temperature and state of charge, respectively. The violation frequency statistics unit calculates the weighted violation frequency. ; Element regulator according to and Dynamic output fusion weights Among them, when and The system is currently in normal operating condition. Neural solvers are preferred; when or The system determines that it is in a limit approximation condition. The symbolic verifier is forced to be used, and the hybrid optimization objective function is: In the formula: For the final optimized scheduling action; This is the loss function for the neural module; For symbolic knowledge base The constraint loss function; These are thermodynamic prior parameters.

7. The cross-seasonal intelligent scheduling method for thermal energy storage based on thermodynamic physical mechanism constraints according to claim 6, characterized in that: The thermodynamic symbol knowledge base It includes three types of core symbolic equations: The first category consists of symbols for the heat balance equation. The expression is: Used to constrain the energy conservation of thermal storage systems; The second category is SOC-defined equation symbols. The expression is: in, This represents the system's maximum thermal storage capacity, used to map temperature to the percentage of thermal storage capacity. The third category consists of symbols for safety and economic dispatch constraint equations. With the goal of minimizing time-of-use electricity costs, the constraints include: in, This is the maximum power adjustment step size allowed by the system.

8. The cross-seasonal intelligent scheduling method for thermal energy storage based on thermodynamic physical mechanism constraints according to claim 1, characterized in that: The method further includes a thermodynamically constraint-aware federated hierarchical collaborative scheduling step, employing a three-layer federated architecture of global coordination layer, regional coordination layer, and local execution layer; the local execution layer uses a proximal policy optimization algorithm with thermodynamic constraint penalties to iterate local policies and prune parameters. Generalized dominance estimation coefficient When the strategy deduction triggers a thermodynamic limit breach, a circuit breaker mechanism is triggered to cut off reward feedback; when the regional coordination layer and the global coordination layer aggregate model parameters, the thermodynamic constraint satisfaction of each local node is used as the core aggregation weight, and the aggregation formula is: In the formula: These are global model parameters; For the region Model parameters; For regional aggregation weights; in, , For the region The number of samples, This represents the degree of satisfaction of thermodynamic constraints.

9. The cross-seasonal intelligent scheduling method for thermal energy storage based on thermodynamic physical mechanism constraints according to claim 1, characterized in that: The method also includes thermodynamic safety monitoring and AI decision-making fault tolerance steps. It collects five core safety parameters of the thermal storage system in real time: average temperature, state of charge (SOC), tank pressure, temperature change rate, and SOC change rate. It sets tiered safety thresholds; when a parameter is within a warning range, power limiting is triggered; when a parameter is within a critical range, the current scheduling command is immediately stopped, a safety protection program is initiated, and the system switches to manual control mode. It also sets up a hardware-level hard constraint protection unit independent of the AI ​​decision chain; when the thermal storage temperature exceeds the hardware safety cut-off threshold... SOC is lower than Or pressure exceeds In some cases, an emergency shutdown action is triggered directly without going through the AI ​​decision-making chain.

10. The cross-seasonal intelligent scheduling method for thermal energy storage based on thermodynamic physical mechanism constraints according to claim 1, characterized in that: In step S7, the constraint boundary of the conservative scheduling strategy is: The safety lower limit is set as follows: The lower limit of safe temperature is set to The upper limit of temperature safety is set to ; The maximum power adjustment step size is tightened to Conservative scheduling strategy to maximize the next time step The expected value is the optimization objective, expressed as: The constraints are .

Images