A multi-objective dynamic optimization method and system for an energy management system based on a large language model and reinforcement learning

Abstract

The present application relates to the technical field of energy management, and is especially a multi-objective dynamic optimization method for an energy management system based on a large language model and reinforcement learning, comprising the following steps: obtaining natural language targets, external information and real-time data; mapping the natural language targets and the external information into optimization configurations through a large language model; generating a multi-objective rolling control strategy based on the optimization configurations and the real-time data; issuing the strategy to an edge after safety checking; collecting actual responses and identifying the sources of deviations between cloud strategies, edge instructions and device responses; and updating model parameters according to the sources of deviations, The present application realizes adaptive mapping of user intentions and external events to optimization parameters through a large language model, realizes multi-objective dynamic trade-off through reinforcement learning, and automatically corrects execution deviations through cloud-edge closed-loop feedback, thereby improving the adaptive operation capability of the energy management system in a multi-objective conflict and uncertain environment.

Description

Technical Field

[0001] This invention relates to the field of energy management technology, specifically to a multi-objective dynamic optimization method and system for energy management systems based on large language models and reinforcement learning. Background Technology

[0002] With the widespread integration of photovoltaics, energy storage, controllable loads, electric vehicles, backup power supplies, and distributed energy assets, energy management systems (EMS) need to simultaneously address multiple objectives in uncertain environments, including cost, carbon emissions, maximum demand, renewable energy integration, equipment lifespan, power supply reliability, market revenue, and operational risks. Traditional energy management systems typically rely on manually configured rules, fixed thresholds, or single-objective optimization models. For example, the system controls energy storage charging and discharging according to preset peak-valley electricity prices, performs peak shaving control according to fixed maximum demand thresholds, or formulates scheduling plans based on simple day-ahead forecasts. While these methods facilitate engineering deployment, they are difficult to adapt to real-world operating scenarios involving conflicting objectives, changes in external events, and dynamic shifts in user preferences.

[0003] Existing energy management systems have the following main limitations:

[0004] Unable to handle natural language goals and policy preferences: Existing EMS relies on engineers to manually set operating parameters, goal weights, and constraint thresholds, and cannot directly understand high-level operating goals and policy preferences described by users in natural language, such as "comfort first" and "minimize maximum demand risk," resulting in a gap between user intent and machine execution.

[0005] Unable to utilize unstructured external information: Existing EMS cannot automatically extract and transform optimization parameters from unstructured external information such as news updates, market announcements, weather warnings, and policy documents. For example, it cannot identify the risk of photovoltaic power output decline from weather warnings or capture demand response opportunities from market announcements, resulting in a lag in the system's response to changes in the external environment.

[0006] Insufficient ability to dynamically balance multiple objectives: Existing EMS generally adopt single-objective (such as minimizing electricity costs) or fixed rule control strategies, making it difficult to dynamically balance multiple conflicting objectives such as cost, carbon emissions, maximum demand, renewable energy consumption, equipment lifespan, and market revenue based on operating status and user preferences.

[0007] Lack of cloud-edge closed-loop adaptive capability: There is a discrepancy between the cloud optimization strategy and the actual response of the field equipment (such as communication delay, equipment aging, prediction error). The existing system cannot automatically identify and correct these discrepancies, requiring manual intervention, which reduces the reliability and economy of the system in long-term operation. Summary of the Invention

[0008] The purpose of this invention is to provide a multi-objective dynamic optimization method and system for energy management systems based on large language models and reinforcement learning, so as to solve the problems mentioned in the background art.

[0009] To achieve the above objectives, the present invention provides the following technical solution:

[0010] A multi-objective dynamic optimization method for energy management systems based on large language models and reinforcement learning includes the following steps:

[0011] S1, acquire data, acquire natural language operation goals, strategy preferences and constraint descriptions input by users or operators, acquire external information related to the energy system, and acquire real-time operation data of the energy system;

[0012] S2, the large language model identifies the target object, priority, hard constraints, soft constraints, time range, triggering conditions, and risk preferences in the natural language operation objectives, strategy preferences, and constraint descriptions. It extracts one or more of energy events, market events, weather events, policy events, fuel events, power grid events, and equipment events from external information. It maps the target object, priority, hard constraints, soft constraints, triggering conditions, risk preferences, and events to objective function terms, constraints, dynamic weights, risk factors, prediction correction factors, reserve capacity requirements, market participation triggering conditions, or equipment operation boundaries. It evaluates the credibility of each parameter mapping result and decides whether to automatically adopt, trigger manual confirmation, or adopt the default configuration based on the evaluation results, generating a machine-readable optimized configuration.

[0013] S3, based on real-time operating data and optimized configuration, generates one or more uncertain scenarios among load, renewable energy output, electricity price, market price, carbon emission factor and equipment availability within a future time window. Based on the uncertain scenarios and optimized configuration, it uses one or more combinations of reinforcement learning model, deep reinforcement learning model, online learning model, model predictive control model, stochastic optimization model, robust optimization model and mixed integer optimization model to generate a multi-objective rolling control strategy within the future time window.

[0014] S4 performs safety constraint verification on the multi-objective rolling control strategy and sends the control strategy that passes the safety constraint verification to the edge controller or field execution unit; the edge controller or field execution unit converts the control strategy into control instructions that can be executed by the field devices and sends the control instructions to one or more field devices among distributed energy devices, energy storage devices, renewable energy inverter devices, backup power devices, controllable load devices, charging devices, power distribution monitoring devices and metering devices.

[0015] S5, collect the actual response and operating status of the field equipment, calculate the power deviation between the planned power generated in the cloud, the command power issued and the actual power of the field equipment, calculate the status deviation between the planned equipment status, the commanded equipment status and the actual equipment status, identify the source of the deviation, and classify the source of the deviation as one or more of the following: communication anomaly, equipment constraint, equipment failure, prediction error, user behavior change, market signal change, external event change and human intervention.

[0016] S6. When the deviation source is equipment constraints or equipment failure, update the equipment operating boundary or equipment availability parameters. When the deviation source is prediction error, update the prediction correction factor or uncertainty scenario. When the deviation source is user behavior change, update the load scenario or user energy consumption constraint. When the deviation source is market signal change or external event change, update the target weight, risk factor, market participation trigger condition or reserve capacity requirement. The update adopts online learning or incremental learning method. Under the premise of maintaining the stability of the existing optimization strategy, the model parameters are gradually adjusted according to the deviation information to avoid control strategy oscillation caused by parameter mutation and enter the next round of rolling optimization.

[0017] As a preferred embodiment of the present invention, the external information in S1 includes one or more of the following: news updates, market dynamics, policy announcements, weather information, electricity price information, ancillary service prices, fuel prices, carbon emission factors, and equipment operation and maintenance information. The real-time operating data in S1 includes one or more of the following: load data, renewable energy output data, energy storage status of charge, equipment availability status, grid connection point power, power distribution equipment capacity, electricity price data, market price data, and historical operating data. The energy system in S1 is an industrial and commercial park energy system, a photovoltaic energy storage system, a microgrid system, a weak grid system, an off-grid energy system, a building energy system, an industrial load energy system, a charging station energy system, an energy storage power station system, a virtual power plant system, or other integrated energy systems composed of multiple types of energy resources.

[0018] As a preferred embodiment of the present invention, the natural language operation objectives, strategy preferences, and constraint descriptions in S2 include one or more of the following: cost objectives, carbon emission objectives, maximum demand objectives, renewable energy consumption objectives, energy storage life objectives, fuel consumption objectives, operational reliability objectives, market revenue objectives, ancillary service participation objectives, comfort objectives, risk preferences, and user energy consumption constraints. The optimization configuration in S2 includes one or more of the following: objective function, constraint conditions, objective weights, risk factors, prediction correction factors, scenario parameters, penalty terms, market participation conditions, and equipment operating boundaries. The optimization configuration uses JSON files, parameter dictionaries, rule templates, optimization model configuration files, tables, vector representations, or combinations thereof. The objective function includes one or more of the following: electricity cost, maximum demand cost, carbon emission cost, energy storage degradation cost, fuel power generation cost, renewable energy curtailment penalty, market revenue, ancillary service revenue, operational risk penalty, reliability penalty, and user comfort penalty. The weights of each objective term in the objective function are dynamically determined by the large language model based on the natural language operation objectives, strategy preferences, constraint descriptions, and external information, and are recalculated by the large language model when external information changes or user preferences are updated. The constraints generated include one or more of the following: energy storage state of charge constraints, energy storage charge and discharge power constraints, energy storage cycle count constraints, grid connection point power constraints, power distribution equipment capacity constraints, backup power output and ramping constraints, renewable energy inverter equipment power constraints, controllable load regulation constraints, charging equipment power constraints, temperature control or HVAC equipment operation constraints, market participation constraints, and user-specified constraints. The constraints also include hard constraints and soft constraints. Hard constraints are constraints that cannot be violated, while soft constraints are constraints that are allowed to be violated within a preset range and are subject to corresponding penalties. The large language model automatically identifies hard constraints and soft constraints based on the constraint descriptions input by the user.

[0019] As a preferred embodiment of the present invention, the uncertain scenario in S3 is generated by a probability prediction model, a scenario tree generation model, a Monte Carlo sampling model, a time series prediction model, a generative model, a large language model-assisted scenario generation model, or a combination thereof. The multi-objective rolling control strategy in S3 includes one or more of the following: energy storage charging and discharging power, energy storage reserve capacity, backup power supply start-stop status and output, controllable load adjustment, charging equipment power, grid connection point power limit, renewable energy consumption strategy, market quotation suggestion, ancillary service capacity reservation, and temperature control or HVAC equipment operation strategy. The multi-objective rolling control strategy is updated at preset time intervals, and the length and update frequency of each rolling time window are determined by the large language model based on user input. The time range description and the event time window in external information are dynamically determined. The state space of the reinforcement learning model or deep reinforcement learning model in S3 includes one or more of the following: current load, predicted load, renewable energy forecast, energy storage state of charge, equipment availability, electricity price, market price, carbon emission factor, user target weight, risk factor, and historical execution deviation. The action space includes one or more of the following: energy storage power, backup power output, controllable load adjustment, charging equipment power, grid connection point power limit, market participation capacity, and reserve capacity. The reward function includes one or more of the following: electricity cost, maximum demand cost, carbon emission cost, energy storage degradation cost, fuel consumption cost, market revenue, risk penalty, and constraint violation penalty.

[0020] As a preferred embodiment of the present invention, the reinforcement learning model or deep reinforcement learning model adopts a secure reinforcement learning framework. During the policy exploration and training process, security constraints are embedded as hard constraints into the action selection or policy update process to ensure that the exploration actions during the training process do not violate the physical boundaries of the equipment and the operation boundaries of the power grid. The reinforcement learning model or deep reinforcement learning model adopts a multi-agent reinforcement learning framework, which treats energy storage equipment, renewable energy equipment, backup power equipment and controllable load equipment as independent agents and performs collaborative optimization through centralized training and decentralized execution or information sharing mechanisms.

[0021] As a preferred embodiment of the present invention, the safety constraint verification in S4 includes multi-level safety verification, specifically: the first level is equipment-level safety verification, checking whether individual devices such as energy storage, inverters, and backup power supplies meet their physical operating boundaries; the second level is system-level safety verification, checking whether system-level constraints such as grid connection point power and distribution capacity are met; the third level is user-level safety verification, checking whether user-specified energy consumption constraints and comfort requirements are met; the fourth level is market-level safety verification, checking whether market rules such as market participation capacity and pricing rules are met; failure at any level of verification triggers strategy correction or rollback of safety constraint verification, and the specific steps of safety constraint verification are as follows: The multi-objective rolling control strategy is checked to ensure it meets the physical boundaries of the equipment, the operation boundaries of the power grid, the energy consumption boundaries of the users, the market rule boundaries, and the preset safety boundaries. If the multi-objective rolling control strategy violates at least one of these boundaries, it is limited, replaced, re-optimized, or reverted to the preset safety strategy. Only control strategies that meet the safety constraints after being checked or corrected are sent to the edge controller or field execution unit. The preset safety strategy includes one or more of the following: maintaining a safe state of charge for energy storage, limiting the power of the grid connection point to not exceed the safety threshold, maintaining power supply to critical loads, stopping unnecessary controllable load adjustments, and keeping the backup power supply in a safe operating range.

[0022] As a preferred embodiment of the present invention, the identification of deviation sources in S5 includes attributing the deviation between cloud strategy and edge instructions to communication anomalies or instruction conversion errors; attributing the deviation between edge instructions and actual device response to device constraints, device failures, or on-site manual intervention; and attributing the deviation between cloud strategy and actual device response, combined with prediction data, to prediction errors, changes in user behavior, changes in market signals, or changes in external events. The update adopts online learning or incremental learning methods, and while maintaining the stability of the existing optimization strategy, the model parameters are gradually adjusted according to the deviation information to avoid control strategy oscillations caused by parameter mutations.

[0023] As a preferred embodiment of the present invention, in the application scenario of industrial and commercial parks, the large language model transforms the user's natural language objectives regarding maximum demand, electricity cost, and renewable energy self-consumption into maximum demand constraints, electricity cost objectives, and power curtailment penalties, and the multi-objective optimization decision module generates energy storage charging and discharging strategies, controllable load adjustment strategies, and grid connection point power control strategies.

[0024] In scenarios where energy storage participates in the electricity market or ancillary services, the large language model transforms market announcements, price information, and user market participation preferences into market participation constraints, revenue parameters, and capacity reservation requirements. The multi-objective optimization decision module then generates control strategies among energy arbitrage, ancillary service revenue, and energy storage degradation costs.

[0025] In weak grid or off-grid photovoltaic-storage-diesel application scenarios, the large language model transforms fuel price, weather information and power supply reliability objectives into fuel power generation cost parameters, renewable energy output correction factors and energy storage backup capacity requirements, and generates a coordinated control strategy for renewable energy, energy storage and backup power by the multi-objective optimization decision module.

[0026] In charging station or building energy application scenarios, the large language model transforms user energy preferences, electricity price information, equipment operation plans and demand response events into charging power constraints, temperature control or HVAC equipment operation constraints, demand response capacity and comfort penalty terms. The multi-objective optimization decision module then generates a collaborative control strategy for charging equipment, energy storage equipment, temperature control or HVAC equipment and controllable loads.

[0027] A multi-objective dynamic optimization system for energy management systems based on large language models and reinforcement learning, including...

[0028] The user intent input module is used to obtain the natural language operation goals, strategy preferences, and constraint descriptions input by users or operators;

[0029] The external information acquisition module is used to acquire one or more of the following information related to the energy system: news updates, market dynamics, policy announcements, weather information, electricity price information, ancillary service prices, fuel prices, carbon emission factors, and equipment operation and maintenance information.

[0030] Run the data acquisition module to obtain real-time operating data of the energy system;

[0031] The large language model parsing module is used to perform semantic parsing, event extraction, parameter mapping, and confidence evaluation on the natural language operation target, policy preference, constraint description, and external information to generate machine-readable optimized configurations.

[0032] The data fusion module is used to fuse the real-time operating data and the optimized configuration;

[0033] The uncertainty modeling module is used to generate one or more uncertainty scenarios among load, renewable energy output, electricity price, market price, carbon emission factor, and equipment availability within a future time window;

[0034] The multi-objective optimization decision module is used to generate multi-objective rolling control strategies within future time windows by employing reinforcement learning models, deep reinforcement learning models, online learning models, model predictive control models, stochastic optimization models, robust optimization models, mixed integer optimization models, or combinations thereof.

[0035] A safety constraint verification module is used to perform safety constraint verification on the multi-objective rolling control strategy;

[0036] The cloud-edge communication module is used to send control policies that have passed security constraint verification to edge controllers or field execution units, and to receive data transmitted back from edge controllers or field execution units.

[0037] An edge execution module is used to convert the control strategy into control commands that can be executed by field devices and to control the field devices;

[0038] The closed-loop feedback correction module is used to compare the deviations between the control strategy generated in the cloud, the control commands issued by the edge controller or field execution unit, and the actual response of the field equipment, calculate the deviations and identify the sources of the deviations, and selectively update the optimized configuration, the uncertainty scenario, the equipment operating boundary, the target weight, or the multi-target rolling control strategy according to the type of the deviation source.

[0039] As a preferred embodiment of the present invention, the large language model parsing module includes an intent recognition unit, an external event extraction unit, a parameter mapping unit, a confidence evaluation unit, and a machine-readable configuration generation unit. The intent recognition unit is configured to: recognize the target object, time range, priority, hard constraints, and soft constraints in the natural language input by the user; the external event extraction unit is configured to: extract energy events, market events, weather events, policy events, fuel events, power grid events, and equipment events from news updates, market updates, policy announcements, weather information, and equipment operation and maintenance information; the parameter mapping unit is configured to: map the intent recognition results and event extraction results to objective function terms, constraints, dynamic weights, risk factors, prediction correction factors, reserve capacity requirements, market participation trigger conditions, or equipment operating boundaries; the confidence evaluation unit is configured to: generate a confidence score for each mapping result; and the machine-readable configuration generation unit is configured to: output the mapping results as JSON, parameter dictionary, rule template, vector representation, table, or optimization model configuration file.

[0040] The security constraint verification module includes a device security verification unit, a power grid security verification unit, a user constraint verification unit, a market rule verification unit, and a strategy rollback unit;

[0041] The edge execution module is configured to execute a preset security policy or the control policy that has most recently passed the security constraint verification when cloud-edge communication is abnormal or cloud control policy is unavailable.

[0042] The closed-loop feedback correction module includes a deviation calculation unit, a deviation source identification unit, and a parameter adaptive update unit. The deviation calculation unit is used to calculate the power deviation and state deviation between cloud strategies, edge commands, and actual device responses. The deviation source identification unit is used to classify deviations into one or more of the following: communication anomalies, device constraints, device failures, prediction errors, changes in user behavior, changes in market signals, changes in external events, and human intervention. The parameter adaptive update unit is used to selectively update device operating boundaries, device availability parameters, prediction correction factors, uncertainty scenarios, target weights, risk factors, market participation trigger conditions, or reserve capacity requirements based on the type of deviation source.

[0043] The data fusion module is configured to integrate the real-time operating data, the optimized configuration, and the historical execution deviation into a unified state representation. The state representation includes the current device state, the predicted state, the risk state, the target weight, the constraint set, and the market parameters, which are then called by the uncertainty modeling module and the multi-objective optimization decision-making module.

[0044] The multi-objective optimization decision module is configured to support the switching and combination of various optimization algorithms, including using a mixed-integer optimization model when explicit processing of discrete decisions is required; using model predictive control when dynamic prediction and constraints are required; using reinforcement learning or deep reinforcement learning when the state space is complex or difficult to model accurately; and using stochastic optimization or robust optimization when dealing with extreme uncertainty. It generates a confidence score between 0 and 1 for each parameter mapping result and sets a first threshold and a second threshold, with the first threshold being greater than the second threshold. When the confidence score is higher than the first threshold, the mapping result is automatically adopted; when the confidence score is between the first and second thresholds, manual confirmation is triggered; and when the confidence score is lower than the second threshold, template rollback or a safe default configuration is used.

[0045] Compared with the prior art, the beneficial effects of the present invention are:

[0046] 1. In this invention, the user's natural language objectives, policy preferences, constraint descriptions, and external unstructured information are transformed into objective functions, constraints, dynamic weights, risk factors, and model parameters in the energy optimization model through a large language model. This bridges the gap between human intent and machine-executable optimization models. Simultaneously, it enables multi-objective dynamic balancing among cost, carbon emissions, maximum demand, equipment lifespan, renewable energy consumption, fuel consumption, market revenue, reliability, and operational risks, avoiding the limitations of traditional single-objective or fixed-rule control strategies in complex scenarios. Furthermore, by utilizing reinforcement learning, deep reinforcement learning, online learning, model predictive control, stochastic optimization, robust optimization, mixed-integer optimization, or combinations thereof, rolling control strategies can be generated for uncertainties such as electricity prices, load, renewable energy output, market prices, and equipment availability.

[0047] 2. In this invention, a closed loop is formed by cloud optimization, edge execution and device feedback, which can identify and correct the deviation between cloud policies, edge commands and actual device responses, thereby improving the system's adaptive operation capability. At the same time, the EMS intelligent optimization kernel can serve as a common platform capability for scenarios such as industrial and commercial parks, microgrids, energy storage stations, charging stations, buildings, industrial loads and virtual power plants.

[0048] 3. In this invention, the safety constraint verification module performs equipment, power grid, user, and market constraint checks before the control strategy is executed, thereby improving the engineering safety of the AI-generated control strategy. Attached Figure Description

[0049] Figure 1 This is a schematic diagram of the overall system architecture of the present invention.

[0050] Figure 2 This is a schematic diagram of the multi-objective dynamic optimization method of the present invention.

[0051] Figure 3 This is a schematic diagram of the information conversion process of the large language model of the present invention.

[0052] Figure 4 This is a schematic diagram of the cloud-edge closed-loop control process of the present invention.

[0053] Figure 5 This is a schematic diagram of the multi-objective optimization decision structure of the present invention. Detailed Implementation

[0054] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0055] To facilitate understanding of the present invention, a more comprehensive description of the invention will be given below with reference to the accompanying drawings, and several embodiments of the invention will be provided. However, the present invention can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided so that the disclosure of the present invention will be more thorough and complete.

[0056] For examples, please refer to Figure 1-5 The present invention provides a technical solution:

[0057] A multi-objective dynamic optimization method for energy management systems based on large language models and reinforcement learning includes the following steps:

[0058] S1, acquire data, acquire natural language operation goals, strategy preferences and constraint descriptions input by users or operators, acquire external information related to the energy system, and acquire real-time operation data of the energy system;

[0059] S2, the large language model identifies the target object, priority, hard constraints, soft constraints, time range, triggering conditions, and risk preferences in the natural language operation objectives, strategy preferences, and constraint descriptions. It extracts one or more of energy events, market events, weather events, policy events, fuel events, power grid events, and equipment events from external information. It maps the target object, priority, hard constraints, soft constraints, triggering conditions, risk preferences, and events to objective function terms, constraints, dynamic weights, risk factors, prediction correction factors, reserve capacity requirements, market participation triggering conditions, or equipment operation boundaries. It evaluates the credibility of each parameter mapping result and decides whether to automatically adopt, trigger manual confirmation, or adopt the default configuration based on the evaluation results, generating a machine-readable optimized configuration.

[0060] S3, based on real-time operating data and optimized configuration, generates one or more uncertain scenarios among load, renewable energy output, electricity price, market price, carbon emission factor and equipment availability within a future time window. Based on the uncertain scenarios and optimized configuration, it uses one or more combinations of reinforcement learning model, deep reinforcement learning model, online learning model, model predictive control model, stochastic optimization model, robust optimization model and mixed integer optimization model to generate a multi-objective rolling control strategy within the future time window.

[0061] S4 performs safety constraint verification on the multi-objective rolling control strategy and sends the control strategy that passes the safety constraint verification to the edge controller or field execution unit; the edge controller or field execution unit converts the control strategy into control instructions that can be executed by the field devices and sends the control instructions to one or more field devices among distributed energy devices, energy storage devices, renewable energy inverter devices, backup power devices, controllable load devices, charging devices, power distribution monitoring devices and metering devices.

[0062] S5, collect the actual response and operating status of the field equipment, calculate the power deviation between the planned power generated in the cloud, the command power issued and the actual power of the field equipment, calculate the status deviation between the planned equipment status, the commanded equipment status and the actual equipment status, identify the source of the deviation, and classify the source of the deviation as one or more of the following: communication anomaly, equipment constraint, equipment failure, prediction error, user behavior change, market signal change, external event change and human intervention.

[0063] S6. When the deviation source is equipment constraints or equipment failure, update the equipment operating boundary or equipment availability parameters. When the deviation source is prediction error, update the prediction correction factor or uncertainty scenario. When the deviation source is user behavior change, update the load scenario or user energy consumption constraint. When the deviation source is market signal change or external event change, update the target weight, risk factor, market participation trigger condition or reserve capacity requirement. The update adopts online learning or incremental learning method. Under the premise of maintaining the stability of the existing optimization strategy, the model parameters are gradually adjusted according to the deviation information to avoid control strategy oscillation caused by parameter mutation and enter the next round of rolling optimization.

[0064] Furthermore, the external information in S1 includes one or more of the following: news updates, market dynamics, policy announcements, weather information, electricity price information, ancillary service prices, fuel prices, carbon emission factors, and equipment operation and maintenance information. The real-time operating data in S1 includes one or more of the following: load data, renewable energy output data, energy storage status of charge, equipment availability status, grid connection point power, power distribution equipment capacity, electricity price data, market price data, and historical operating data. The energy system in S1 is an industrial and commercial park energy system, a photovoltaic energy storage system, a microgrid system, a weak grid system, an off-grid energy system, a building energy system, an industrial load energy system, a charging station energy system, an energy storage power station system, a virtual power plant system, or other integrated energy systems composed of multiple types of energy resources.

[0065] Furthermore, the natural language operation objectives, strategy preferences, and constraint descriptions in S2 include one or more of the following: cost objectives, carbon emission objectives, maximum demand objectives, renewable energy consumption objectives, energy storage life objectives, fuel consumption objectives, operational reliability objectives, market revenue objectives, ancillary service participation objectives, comfort objectives, risk preferences, and user energy consumption constraints. The optimization configuration in S2 includes one or more of the following: objective function, constraints, objective weights, risk factors, prediction correction factors, scenario parameters, penalty terms, market participation conditions, and equipment operating boundaries. The optimization configuration uses JSON files, parameter dictionaries, rule templates, optimization model configuration files, tables, vector representations, or combinations thereof. The objective function includes one or more of the following: electricity cost, maximum demand cost, carbon emission cost, energy storage degradation cost, fuel power generation cost, renewable energy curtailment penalty, market revenue, ancillary service revenue, operational risk penalty, reliability penalty, and user comfort penalty. The weights of each objective term in the objective function are dynamically determined by the large language model based on the natural language operation objectives, strategy preferences, constraint descriptions, and external information, and are recalculated by the large language model when external information changes or user preferences are updated. The constraints generated include one or more of the following: energy storage state of charge constraints, energy storage charge and discharge power constraints, energy storage cycle count constraints, grid connection point power constraints, power distribution equipment capacity constraints, backup power output and ramping constraints, renewable energy inverter equipment power constraints, controllable load regulation constraints, charging equipment power constraints, temperature control or HVAC equipment operation constraints, market participation constraints, and user-specified constraints. The constraints also include hard constraints and soft constraints. Hard constraints are constraints that cannot be violated, while soft constraints are constraints that are allowed to be violated within a preset range and are subject to corresponding penalties. The large language model automatically identifies hard constraints and soft constraints based on the constraint descriptions input by the user.

[0066] Furthermore, in S3, the uncertain scenarios are generated through a probability prediction model, a scenario tree generation model, a Monte Carlo sampling model, a time series prediction model, a generative model, a large language model-assisted scenario generation model, or a combination thereof. The multi-objective rolling control strategy in S3 includes one or more of the following: energy storage charging and discharging power, energy storage reserve capacity, backup power supply start / stop status and output, controllable load adjustment, charging equipment power, grid connection point power limit, renewable energy consumption strategy, market pricing suggestions, ancillary service capacity reservation, and temperature control or HVAC equipment operation strategy. The multi-objective rolling control strategy is updated at preset time intervals, and the length and update frequency of each rolling time window are determined by the large language model based on the user-input time range. The event time window in the surrounding description and external information is dynamically determined. The state space of the reinforcement learning model or deep reinforcement learning model in S3 includes one or more of the following: current load, predicted load, renewable energy forecast, energy storage state of charge, equipment availability, electricity price, market price, carbon emission factor, user target weight, risk factor, and historical execution deviation. The action space includes one or more of the following: energy storage power, backup power output, controllable load adjustment, charging equipment power, grid connection point power limit, market participation capacity, and reserve capacity. The reward function includes one or more of the following: electricity cost, maximum demand cost, carbon emission cost, energy storage degradation cost, fuel consumption cost, market revenue, risk penalty, and constraint violation penalty.

[0067] Furthermore, the reinforcement learning model or deep reinforcement learning model adopts a safe reinforcement learning framework, embedding safety constraints as hard constraints into the action selection or policy update process during policy exploration and training to ensure that the exploration actions during training do not violate the physical boundaries of the equipment and the operation boundaries of the power grid. The reinforcement learning model or deep reinforcement learning model adopts a multi-agent reinforcement learning framework, treating energy storage equipment, renewable energy equipment, backup power equipment and controllable load equipment as independent agents, and performing collaborative optimization through centralized training and decentralized execution or information sharing mechanisms.

[0068] Furthermore, the safety constraint verification in S4 includes multi-level safety verification, specifically: the first level is device-level safety verification, checking whether individual devices such as energy storage, inverters, and backup power supplies meet their physical operating boundaries; the second level is system-level safety verification, checking whether system-level constraints such as grid connection point power and distribution capacity are met; the third level is user-level safety verification, checking whether user-specified energy consumption constraints and comfort requirements are met; the fourth level is market-level safety verification, checking whether market rules such as market participation capacity and pricing rules are met; failure at any level of verification triggers a strategy correction or rollback of the safety constraint verification. The specific steps of the safety constraint verification are: checking... Whether the multi-objective rolling control strategy meets the physical boundaries of the equipment, the operation boundaries of the power grid, the energy consumption boundaries of the users, the market rule boundaries, and the preset safety boundaries. When the multi-objective rolling control strategy violates at least one of the boundaries, the multi-objective rolling control strategy is limited, replaced, re-optimized, or reverted to the preset safety strategy. Only the control strategy that meets the safety constraints after inspection or correction is issued to the edge controller or field execution unit. The preset safety strategy includes one or more of the following: maintaining a safe state of charge of energy storage, limiting the power of the grid connection point to not exceed the safety threshold, maintaining power supply to critical loads, stopping unnecessary controllable load adjustments, and keeping the backup power supply in the safe operating range.

[0069] Furthermore, in S5, identifying the sources of deviation includes attributing the deviation between the cloud strategy and the edge command to communication anomalies or command conversion errors; attributing the deviation between the edge command and the actual device response to device constraints, device failures, or on-site manual intervention; and attributing the deviation between the cloud strategy and the actual device response, combined with prediction data, to prediction errors, changes in user behavior, changes in market signals, or changes in external events. The update adopts online learning or incremental learning methods, and while maintaining the stability of the existing optimization strategy, it gradually adjusts the model parameters based on the deviation information to avoid control strategy oscillations caused by parameter mutations.

[0070] Furthermore, in the application scenario of industrial and commercial parks, the large language model transforms the user's natural language objectives regarding maximum demand, electricity costs, and renewable energy self-consumption into maximum demand constraints, electricity cost objectives, and power curtailment penalties. The multi-objective optimization decision module then generates energy storage charging and discharging strategies, controllable load adjustment strategies, and grid connection point power control strategies.

[0071] In scenarios where energy storage participates in the electricity market or ancillary services, the large language model transforms market announcements, price information, and user market participation preferences into market participation constraints, revenue parameters, and capacity reservation requirements. The multi-objective optimization decision module then generates control strategies among energy arbitrage, ancillary service revenue, and energy storage degradation costs.

[0072] In weak grid or off-grid photovoltaic-storage-diesel application scenarios, the large language model transforms fuel price, weather information and power supply reliability objectives into fuel power generation cost parameters, renewable energy output correction factors and energy storage backup capacity requirements, and generates a coordinated control strategy for renewable energy, energy storage and backup power by the multi-objective optimization decision module.

[0073] In charging station or building energy application scenarios, the large language model transforms user energy preferences, electricity price information, equipment operation plans and demand response events into charging power constraints, temperature control or HVAC equipment operation constraints, demand response capacity and comfort penalty terms. The multi-objective optimization decision module then generates a collaborative control strategy for charging equipment, energy storage equipment, temperature control or HVAC equipment and controllable loads.

[0074] A multi-objective dynamic optimization system for energy management systems based on large language models and reinforcement learning, including...

[0075] The user intent input module is used to obtain the natural language operation goals, strategy preferences, and constraint descriptions input by users or operators;

[0076] The external information acquisition module is used to acquire one or more of the following information related to the energy system: news updates, market dynamics, policy announcements, weather information, electricity price information, ancillary service prices, fuel prices, carbon emission factors, and equipment operation and maintenance information.

[0077] Run the data acquisition module to obtain real-time operating data of the energy system;

[0078] The large language model parsing module is used to perform semantic parsing, event extraction, parameter mapping, and confidence evaluation on the natural language operation target, policy preference, constraint description, and external information to generate machine-readable optimized configurations.

[0079] The data fusion module is used to fuse the real-time operating data and the optimized configuration;

[0080] The uncertainty modeling module is used to generate one or more uncertainty scenarios among load, renewable energy output, electricity price, market price, carbon emission factor, and equipment availability within a future time window;

[0081] The multi-objective optimization decision module is used to generate multi-objective rolling control strategies within future time windows by employing reinforcement learning models, deep reinforcement learning models, online learning models, model predictive control models, stochastic optimization models, robust optimization models, mixed integer optimization models, or combinations thereof.

[0082] A safety constraint verification module is used to perform safety constraint verification on the multi-objective rolling control strategy;

[0083] The cloud-edge communication module is used to send control policies that have passed security constraint verification to edge controllers or field execution units, and to receive data transmitted back from edge controllers or field execution units.

[0084] An edge execution module is used to convert the control strategy into control commands that can be executed by field devices and to control the field devices;

[0085] The closed-loop feedback correction module is used to compare the deviations between the control strategy generated in the cloud, the control commands issued by the edge controller or field execution unit, and the actual response of the field equipment, calculate the deviations and identify the sources of the deviations, and selectively update the optimized configuration, the uncertainty scenario, the equipment operating boundary, the target weight, or the multi-target rolling control strategy according to the type of the deviation source.

[0086] Furthermore, the large language model parsing module includes an intent recognition unit, an external event extraction unit, a parameter mapping unit, a confidence evaluation unit, and a machine-readable configuration generation unit. The intent recognition unit is configured to: identify the target object, time range, priority, hard constraints, and soft constraints in the natural language input by the user; the external event extraction unit is configured to: extract energy events, market events, weather events, policy events, fuel events, power grid events, and equipment events from news updates, market updates, policy announcements, weather information, and equipment operation and maintenance information; the parameter mapping unit is configured to: map the intent recognition results and event extraction results to objective function terms, constraints, dynamic weights, risk factors, prediction correction factors, reserve capacity requirements, market participation trigger conditions, or equipment operating boundaries; the confidence evaluation unit is configured to: generate a confidence score for each mapping result; and the machine-readable configuration generation unit is configured to: output the mapping results as JSON, parameter dictionary, rule template, vector representation, table, or optimization model configuration file.

[0087] The security constraint verification module includes a device security verification unit, a power grid security verification unit, a user constraint verification unit, a market rule verification unit, and a strategy rollback unit;

[0088] The edge execution module is configured to execute a preset security policy or the control policy that has most recently passed the security constraint verification when cloud-edge communication is abnormal or cloud control policy is unavailable.

[0089] The closed-loop feedback correction module includes a deviation calculation unit, a deviation source identification unit, and a parameter adaptive update unit. The deviation calculation unit is used to calculate the power deviation and state deviation between cloud strategies, edge commands, and actual device responses. The deviation source identification unit is used to classify deviations into one or more of the following: communication anomalies, device constraints, device failures, prediction errors, changes in user behavior, changes in market signals, changes in external events, and human intervention. The parameter adaptive update unit is used to selectively update device operating boundaries, device availability parameters, prediction correction factors, uncertainty scenarios, target weights, risk factors, market participation trigger conditions, or reserve capacity requirements based on the type of deviation source.

[0090] The data fusion module is configured to integrate the real-time operating data, the optimized configuration, and the historical execution deviation into a unified state representation. The state representation includes the current device state, the predicted state, the risk state, the target weight, the constraint set, and the market parameters, which are then called by the uncertainty modeling module and the multi-objective optimization decision-making module.

[0091] The multi-objective optimization decision module is configured to support the switching and combination of various optimization algorithms, including using a mixed-integer optimization model when explicit processing of discrete decisions is required; using model predictive control when dynamic prediction and constraints are required; using reinforcement learning or deep reinforcement learning when the state space is complex or difficult to model accurately; and using stochastic optimization or robust optimization when dealing with extreme uncertainty. The module evaluates the credibility of each parameter mapping result by generating a confidence score between 0 and 1 for each parameter mapping result and setting a first threshold and a second threshold, where the first threshold is greater than the second threshold. When the confidence score is higher than the first threshold, the mapping result is automatically adopted; when the confidence score is between the first and second thresholds, manual confirmation is triggered; and when the confidence score is lower than the second threshold, template rollback or a safe default configuration is used.

[0092] Specific Implementation Cases

[0093] Multi-objective dynamic optimization of energy management in typical industrial and commercial parks

[0094] This embodiment uses a typical industrial and commercial park's integrated energy system as an example to illustrate the specific implementation of the present invention. The main energy equipment configuration of the park is as follows: the photovoltaic power generation system has an installed capacity of 1.5 MW; the energy storage system is configured with 2 MW / 4 MWh lithium iron phosphate batteries, with an allowable charge / discharge rate of 0.5C and a safe state of charge (SOC) range set between 10% and 90%; the park's conventional peak load is approximately 3 MW, including lighting, office equipment, and uninterrupted production lines; the controllable load includes the central air conditioning system and some flexible production equipment, with a total regulation capacity of approximately 500 kW; the grid connection point transformer has a contracted capacity of 3.5 MVA; the park implements a peak-valley time-of-use pricing mechanism and can participate in the regional electricity market demand response;

[0095] Users or operators can input operational goals via natural language, such as: "An important client is visiting next Monday, so the office air conditioning must remain on; comfort is paramount. Also, the peak-valley electricity price difference is significant recently, so try to utilize energy storage arbitrage as much as possible, but don't let the maximum demand exceed 3.2 MW. If there is demand response tomorrow afternoon and the price is favorable, we will participate."

[0096] Meanwhile, the external information acquisition module obtained the following external information: the weather forecast showed that "there will be severe convective weather next Monday afternoon, and the output of photovoltaic power is expected to decrease by 40%"; the market announcement showed that "peak shaving demand response will be carried out from 14:00 to 16:00 tomorrow, and the compensation price is 3 yuan / kWh".

[0097] The large language model parsing module performs semantic parsing, event extraction, and parameter mapping on the above natural language objectives and external information to generate machine-readable optimized configurations: Regarding the objective function weights, the electricity cost weight is set to 0.4, the maximum demand cost weight is set to 0.3, the comfort penalty weight is set to 0.8 (significantly higher than the default value due to the user's explicit expression of "comfort priority"), and the market revenue weight is set to 0.6; Regarding hard constraints, the maximum demand limit for grid-connected points is set to 3.2MW, and the office area air conditioning system is marked as an uninterruptible load; Regarding soft constraints, the daily cycle count for energy storage is limited to 2 times; Regarding market participation, the demand response trigger condition is set to a compensation price of no less than 2 yuan / kWh, and the participation time window is from 14:00 to 16:00 the following day; Regarding forecast correction, the photovoltaic output forecast correction factor for next Monday afternoon is set to 0.6.

[0098] The large language model generates a confidence score for each mapping result: for explicit numerical constraints (such as "maximum demand not exceeding 3.2 megawatts"), the system automatically adopts a confidence score higher than the first threshold; for parameters that need to be inferred (such as the specific value of comfort weight), the system triggers a manual confirmation prompt if the confidence score is between the first and second thresholds; for parameters that cannot be inferred from the input, the system adopts the default safety configuration.

[0099] Based on the above machine-readable optimized configuration, the multi-objective optimization decision module adopts deep reinforcement learning (DRL) to generate a rolling control strategy, and the specific design is as follows.

[0100] The state space has 12 dimensions, including: the current time (normalized to between 0 and 1), the predicted output of photovoltaic power, the predicted value of conventional load, the current state of charge of energy storage, the real-time electricity price, the market price of ancillary services, the maximum demand recorded in the current billing cycle, and five dynamic parameters generated by mapping from the large language model: electricity cost weight, demand cost weight, comfort penalty weight, market revenue weight, and photovoltaic output prediction correction factor.

[0101] The action space has three dimensions, including: energy storage charging and discharging power command (a continuous variable, ranging from -2 MW to 2MW, with negative values ​​indicating charging and positive values ​​indicating discharging), controllable load adjustment (a continuous variable, ranging from 0 to 500 kW), and demand response market declared capacity (a continuous variable, ranging from 0 to 500 kW).

[0102] The reward function comprehensively considers multiple objectives and is expressed as: R = -(w1·C_energy + w2·C_demand + w3·P_comfort) + w4·R_market -Penalty_constraint. Where C_energy is the energy cost at the current time step, C_demand is the contribution of the current time step to the maximum demand record, P_comfort is the penalty for the impact of controllable load adjustment on user comfort, R_market is the market participation reward, Penalty_constraint is the maximum negative penalty imposed when hard constraints are violated, and w1 to w4 are weights dynamically assigned by the large language model, which in this embodiment are taken as 0.4, 0.3, 0.8, and 0.6, respectively.

[0103] The reinforcement learning model uses a rolling step size of 15 minutes and a prediction time domain of 24 hours, regenerating the control strategy at each rolling time step. Before the strategy is issued, the safety constraint verification module checks whether the energy storage charge status is within the range of 10% to 90%, whether the grid connection point power does not exceed 3.2 MW, and whether the office area air conditioning is in operation. If any constraint is not met, the strategy is corrected or rolled back.

[0104] In the aforementioned industrial and commercial park scenario, a one-month simulation comparison was conducted between the present invention and the traditional control method. The results are shown in the table below:

[0105] The above comparison results show that the present invention achieves automatic parameterized mapping of user intent and external environmental events through a large language model, and realizes dynamic trade-offs of multiple objectives through deep reinforcement learning. It is superior to traditional control methods in terms of comprehensive economy, demand control, market participation and user intent response, and can adapt to changes in user objectives and external environment without engineers having to manually modify the underlying parameters.

[0106] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A multi-objective dynamic optimization method for energy management systems based on large language models and reinforcement learning, characterized in that, Includes the following steps: S1, acquire data, acquire natural language operation goals, strategy preferences and constraint descriptions input by users or operators, acquire external information related to the energy system, and acquire real-time operation data of the energy system; S2, the large language model identifies the target object, priority, hard constraints, soft constraints, time range, triggering conditions, and risk preferences in the natural language operation objectives, strategy preferences, and constraint descriptions. It extracts one or more of energy events, market events, weather events, policy events, fuel events, power grid events, and equipment events from external information. It maps the target object, priority, hard constraints, soft constraints, triggering conditions, risk preferences, and events to objective function terms, constraints, dynamic weights, risk factors, prediction correction factors, reserve capacity requirements, market participation triggering conditions, or equipment operation boundaries. It evaluates the credibility of each parameter mapping result and decides whether to automatically adopt, trigger manual confirmation, or adopt the default configuration based on the evaluation results, generating a machine-readable optimized configuration. S3, based on real-time operating data and optimized configuration, generates one or more uncertain scenarios among load, renewable energy output, electricity price, market price, carbon emission factor and equipment availability within a future time window. Based on the uncertain scenarios and optimized configuration, it uses one or more combinations of reinforcement learning model, deep reinforcement learning model, online learning model, model predictive control model, stochastic optimization model, robust optimization model and mixed integer optimization model to generate a multi-objective rolling control strategy within the future time window. S4 performs safety constraint verification on the multi-objective rolling control strategy and sends the control strategy that passes the safety constraint verification to the edge controller or field execution unit; the edge controller or field execution unit converts the control strategy into control instructions that can be executed by the field devices and sends the control instructions to one or more field devices among distributed energy devices, energy storage devices, renewable energy inverter devices, backup power devices, controllable load devices, charging devices, power distribution monitoring devices and metering devices. S5, collect the actual response and operating status of the field equipment, calculate the power deviation between the planned power generated in the cloud, the command power issued and the actual power of the field equipment, calculate the status deviation between the planned equipment status, the commanded equipment status and the actual equipment status, identify the source of the deviation, and classify the source of the deviation as one or more of the following: communication anomaly, equipment constraint, equipment failure, prediction error, user behavior change, market signal change, external event change and human intervention. S6. When the deviation source is equipment constraints or equipment failure, update the equipment operating boundary or equipment availability parameters. When the deviation source is prediction error, update the prediction correction factor or uncertainty scenario. When the deviation source is user behavior change, update the load scenario or user energy consumption constraint. When the deviation source is market signal change or external event change, update the target weight, risk factor, market participation trigger condition or reserve capacity requirement. The update adopts online learning or incremental learning method. Under the premise of maintaining the stability of the existing optimization strategy, the model parameters are gradually adjusted according to the deviation information to avoid control strategy oscillation caused by parameter mutation and enter the next round of rolling optimization.

2. The multi-objective dynamic optimization method for energy management systems based on large language models and reinforcement learning according to claim 1, characterized in that: The external information in S1 includes one or more of the following: news updates, market dynamics, policy announcements, weather information, electricity price information, ancillary service prices, fuel prices, carbon emission factors, and equipment operation and maintenance information. The real-time operating data in S1 includes one or more of the following: load data, renewable energy output data, energy storage status of charge, equipment availability status, grid connection point power, power distribution equipment capacity, electricity price data, market price data, and historical operating data. The energy system in S1 is an industrial and commercial park energy system, a photovoltaic energy storage system, a microgrid system, a weak grid system, an off-grid energy system, a building energy system, an industrial load energy system, a charging station energy system, an energy storage power station system, a virtual power plant system, or other integrated energy systems composed of multiple types of energy resources.

3. The multi-objective dynamic optimization method for energy management systems based on large language models and reinforcement learning according to claim 1, characterized in that: The natural language operation objectives, strategy preferences, and constraint descriptions in S2 include one or more of the following: cost objectives, carbon emission objectives, maximum demand objectives, renewable energy consumption objectives, energy storage life objectives, fuel consumption objectives, operational reliability objectives, market revenue objectives, ancillary service participation objectives, comfort objectives, risk preferences, and user energy consumption constraints. The optimization configuration in S2 includes one or more of the following: objective function, constraints, objective weights, risk factors, prediction correction factors, scenario parameters, penalty terms, market participation conditions, and equipment operating boundaries. The optimization configuration uses JSON files, parameter dictionaries, rule templates, optimization model configuration files, tables, vector representations, or combinations thereof. The objective function includes the following: electricity cost, maximum demand cost, carbon emission cost, energy storage degradation cost, fuel power generation cost, renewable energy curtailment penalty, market revenue, ancillary service revenue, operational risk penalty, and reliability penalty. The objective function includes one or more of the following: a target item and a user comfort penalty item. The weights of each objective item in the objective function are dynamically determined by the large language model based on the natural language operation objective, strategy preferences, constraint descriptions, and external information. The weights are regenerated by the large language model when external information changes or user preferences are updated. The constraints include one or more of the following: energy storage state of charge constraints, energy storage charging and discharging power constraints, energy storage cycle number constraints, grid connection point power constraints, power distribution equipment capacity constraints, backup power output and ramping constraints, renewable energy inverter equipment power constraints, controllable load regulation constraints, charging equipment power constraints, temperature control or HVAC equipment operation constraints, market participation constraints, and user-specified constraints. The constraints also include hard constraints and soft constraints. Hard constraints are constraints that cannot be violated, while soft constraints are constraints that are allowed to be violated within a preset range and are subject to corresponding penalties. The large language model automatically identifies hard constraints and soft constraints based on the constraint descriptions input by the user.

4. The multi-objective dynamic optimization method for energy management systems based on large language models and reinforcement learning according to claim 1, characterized in that: In S3, the uncertain scenarios are generated through a probability prediction model, a scenario tree generation model, a Monte Carlo sampling model, a time series prediction model, a generative model, a large language model-assisted scenario generation model, or a combination thereof. The multi-objective rolling control strategy in S3 includes one or more of the following: energy storage charging and discharging power, energy storage reserve capacity, backup power supply start-stop status and output, controllable load adjustment, charging equipment power, grid connection point power limit, renewable energy consumption strategy, market quotation suggestion, ancillary service capacity reservation, and temperature control or HVAC equipment operation strategy. The multi-objective rolling control strategy is updated at a preset time interval. The length and update frequency of each rolling time window are dynamically determined by the large language model based on the user-input time range description and the event time window in external information. The state space of the reinforcement learning model or deep reinforcement learning model in S3 includes one or more of the following: current load, predicted load, renewable energy prediction, energy storage state of charge, equipment availability, electricity price, market price, carbon emission factor, user target weight, risk factor, and historical execution deviation. The action space includes one or more of the following: energy storage capacity, backup power output, controllable load adjustment, charging equipment power, grid connection point power limit, market participation capacity, and reserve capacity; the reward function includes one or more of the following: electricity cost, maximum demand cost, carbon emission cost, energy storage degradation cost, fuel consumption cost, market revenue, risk penalty, and constraint violation penalty.

5. The multi-objective dynamic optimization method for an energy management system based on a large language model and reinforcement learning as described in claim 4, characterized in that: The reinforcement learning model or deep reinforcement learning model adopts a safe reinforcement learning framework, incorporating safety constraints as hard constraints into the action selection or policy update process during policy exploration and training to ensure that the exploration actions during training do not violate the physical boundaries of the equipment and the operation boundaries of the power grid. The reinforcement learning model or deep reinforcement learning model adopts a multi-agent reinforcement learning framework, treating energy storage equipment, renewable energy equipment, backup power equipment and controllable load equipment as independent agents, and performing collaborative optimization through centralized training and decentralized execution or information sharing mechanisms.

6. The multi-objective dynamic optimization method for an energy management system based on a large language model and reinforcement learning as described in claim 1, characterized in that: The safety constraint verification in S4 includes multi-level safety verification, specifically: the first level is device-level safety verification, checking whether individual devices such as energy storage, inverters, and backup power supplies meet their physical operating boundaries; the second level is system-level safety verification, checking whether system-level constraints such as grid connection point power and distribution capacity are met; the third level is user-level safety verification, checking whether user-specified energy consumption constraints and comfort requirements are met; and the fourth level is market-level safety verification, checking whether market rules such as market participation capacity and pricing rules are met. Failure at any level triggers a strategy correction or rollback of the safety constraint verification. The specific steps of the safety constraint verification are: checking the multi-level safety constraint verification... Whether the target rolling control strategy meets the physical boundaries of the equipment, the operation boundaries of the power grid, the energy consumption boundaries of the user, the market rule boundaries, and the preset safety boundaries. When the multi-objective rolling control strategy violates at least one of the boundaries, the multi-objective rolling control strategy is limited, replaced, re-optimized, or reverted to the preset safety strategy. Only the control strategy that meets the safety constraints after inspection or correction is issued to the edge controller or field execution unit. The preset safety strategy includes one or more of the following: maintaining a safe state of charge of energy storage, limiting the power of the grid connection point to not exceed the safety threshold, maintaining the power supply of critical loads, stopping unnecessary controllable load adjustment, and keeping the backup power supply in the safe operating range.

7. The multi-objective dynamic optimization method for energy management systems based on large language models and reinforcement learning according to claim 1, characterized in that: The source of deviation identified in S5 includes attributing the deviation between cloud strategy and edge command to communication anomalies or command conversion errors; attributing the deviation between edge command and actual device response to device constraints, device failure, or on-site manual intervention; and attributing the deviation between cloud strategy and actual device response, combined with prediction data, to prediction errors, changes in user behavior, changes in market signals, or changes in external events. The update adopts online learning or incremental learning methods, and while maintaining the stability of the existing optimization strategy, it gradually adjusts the model parameters according to the deviation information to avoid control strategy oscillations caused by parameter mutations.

8. The multi-objective dynamic optimization method for energy management systems based on large language models and reinforcement learning according to claim 1, characterized in that: In the application scenario of industrial and commercial parks, the large language model transforms the user's natural language objectives regarding maximum demand, electricity cost, and renewable energy self-consumption into maximum demand constraints, electricity cost objectives, and power curtailment penalties. The multi-objective optimization decision module then generates energy storage charging and discharging strategies, controllable load adjustment strategies, and grid connection point power control strategies. In scenarios where energy storage participates in the electricity market or ancillary services, the large language model transforms market announcements, price information, and user market participation preferences into market participation constraints, revenue parameters, and capacity reservation requirements. The multi-objective optimization decision module then generates control strategies among energy arbitrage, ancillary service revenue, and energy storage degradation costs. In weak grid or off-grid photovoltaic-storage-diesel application scenarios, the large language model transforms fuel price, weather information and power supply reliability objectives into fuel power generation cost parameters, renewable energy output correction factors and energy storage backup capacity requirements, and generates a coordinated control strategy for renewable energy, energy storage and backup power by the multi-objective optimization decision module. In charging station or building energy application scenarios, the large language model transforms user energy preferences, electricity price information, equipment operation plans and demand response events into charging power constraints, temperature control or HVAC equipment operation constraints, demand response capacity and comfort penalty terms. The multi-objective optimization decision module then generates a collaborative control strategy for charging equipment, energy storage equipment, temperature control or HVAC equipment and controllable loads.

9. A multi-objective dynamic optimization system for energy management systems based on large language models and reinforcement learning, characterized in that: include The user intent input module is used to obtain the natural language operation goals, strategy preferences, and constraint descriptions input by users or operators; The external information acquisition module is used to acquire one or more of the following information related to the energy system: news updates, market dynamics, policy announcements, weather information, electricity price information, ancillary service prices, fuel prices, carbon emission factors, and equipment operation and maintenance information. Run the data acquisition module to obtain real-time operating data of the energy system; The large language model parsing module is used to perform semantic parsing, event extraction, parameter mapping, and confidence evaluation on the natural language operation target, policy preference, constraint description, and external information to generate machine-readable optimized configurations. The data fusion module is used to fuse the real-time operating data and the optimized configuration; The uncertainty modeling module is used to generate one or more uncertainty scenarios among load, renewable energy output, electricity price, market price, carbon emission factor, and equipment availability within a future time window; The multi-objective optimization decision module is used to generate multi-objective rolling control strategies within future time windows by employing reinforcement learning models, deep reinforcement learning models, online learning models, model predictive control models, stochastic optimization models, robust optimization models, mixed integer optimization models, or combinations thereof. A safety constraint verification module is used to perform safety constraint verification on the multi-objective rolling control strategy; The cloud-edge communication module is used to send control policies that have passed security constraint verification to edge controllers or field execution units, and to receive data transmitted back from edge controllers or field execution units. An edge execution module is used to convert the control strategy into control commands that can be executed by field devices and to control the field devices; The closed-loop feedback correction module is used to compare the deviations between the control strategy generated in the cloud, the control commands issued by the edge controller or field execution unit, and the actual response of the field equipment, calculate the deviations and identify the sources of the deviations, and selectively update the optimized configuration, the uncertainty scenario, the equipment operating boundary, the target weight, or the multi-target rolling control strategy according to the type of the deviation source.

10. A multi-objective dynamic optimization system for an energy management system based on a large language model and reinforcement learning as described in claim 9, characterized in that: The large language model parsing module includes an intent recognition unit, an external event extraction unit, a parameter mapping unit, a confidence evaluation unit, and a machine-readable configuration generation unit. The intent recognition unit is configured to: recognize the target object, time range, priority, hard constraints, and soft constraints in the natural language input by the user; the external event extraction unit is configured to: extract energy events, market events, weather events, policy events, fuel events, power grid events, and equipment events from news updates, market updates, policy announcements, weather information, and equipment operation and maintenance information. The parameter mapping unit is configured to map the intent recognition results and event extraction results to objective function terms, constraints, dynamic weights, risk factors, prediction correction factors, standby capacity requirements, market participation trigger conditions, or equipment operating boundaries; the confidence evaluation unit is configured to generate a confidence score for each mapping result; the machine-readable configuration generation unit is configured to output the mapping results as JSON, parameter dictionary, rule template, vector representation, table, or optimization model configuration file. The security constraint verification module includes a device security verification unit, a power grid security verification unit, a user constraint verification unit, a market rule verification unit, and a strategy rollback unit; The edge execution module is configured to execute a preset security policy or the control policy that has most recently passed the security constraint verification when cloud-edge communication is abnormal or cloud control policy is unavailable. The closed-loop feedback correction module includes a deviation calculation unit, a deviation source identification unit, and a parameter adaptive update unit. The deviation calculation unit is used to calculate the power deviation and state deviation between cloud strategies, edge commands, and actual device responses. The deviation source identification unit is used to classify deviations into one or more of the following: communication anomalies, device constraints, device failures, prediction errors, changes in user behavior, changes in market signals, changes in external events, and human intervention. The parameter adaptive update unit is used to selectively update device operating boundaries, device availability parameters, prediction correction factors, uncertainty scenarios, target weights, risk factors, market participation trigger conditions, or reserve capacity requirements based on the type of deviation source. The data fusion module is configured to integrate the real-time operating data, the optimized configuration, and the historical execution deviation into a unified state representation. The state representation includes the current device state, the predicted state, the risk state, the target weight, the constraint set, and the market parameters, which are then called by the uncertainty modeling module and the multi-objective optimization decision-making module. The multi-objective optimization decision module is configured to support the switching and combination of various optimization algorithms, including using a mixed-integer optimization model when explicit processing of discrete decisions is required; using model predictive control when dynamic prediction and constraints are required; using reinforcement learning or deep reinforcement learning when the state space is complex or difficult to model accurately; and using stochastic optimization or robust optimization when dealing with extreme uncertainty. The module evaluates the credibility of each parameter mapping result by generating a confidence score between 0 and 1 for each parameter mapping result and setting a first threshold and a second threshold, where the first threshold is greater than the second threshold. When the confidence score is higher than the first threshold, the mapping result is automatically adopted; when the confidence score is between the first and second thresholds, manual confirmation is triggered; and when the confidence score is lower than the second threshold, template rollback or a safe default configuration is used.

Images