A hybrid energy storage scheduling method and system based on LLM agent and game

Abstract

The application discloses a kind of hybrid energy storage scheduling method and system based on LLM agent and game. The method comprises: the multiple-source heterogeneous data collected is preprocessed and total power instruction is generated;Response natural language scheduling instruction generates scheduling parameter set including quantization constraint and target weight;Master-slave game iteration is executed by leader agent and multiple follower agents, and basic charge-discharge power curve is obtained;Total power instruction is decomposed into frequency band and fused with basic charge-discharge power curve to generate target scheduling instruction sequence;Target scheduling instruction sequence is issued to the controller of each energy storage device, so that each controller controls corresponding energy storage device to execute charge-discharge operation according to target scheduling instruction sequence;Since the application is decomposed by heterogeneous data and master-slave game equilibrium, the overall flatness cost of energy storage system and comprehensive power consumption cost are significantly reduced, and the adaptability and synergy of hybrid energy storage scheduling are improved through master-slave game between leader and follower.

Description

Technical Field

[0001] This application relates to the field of energy storage scheduling technology, and in particular to a hybrid energy storage scheduling method and system based on LLM agents and game theory. Background Technology

[0002] With the large-scale grid connection of renewable energy sources (photovoltaics and wind power), the demand for energy storage technology in the power system is becoming increasingly urgent. Energy storage systems can smooth out fluctuations in renewable energy, achieve peak shaving and valley filling, and participate in ancillary services. Currently, mainstream energy storage technologies include lithium battery energy storage, flow battery energy storage, flywheel energy storage, supercapacitors, compressed air energy storage, hydrogen energy storage, and molten salt thermal energy storage. Different energy storage technologies vary significantly in terms of response time, energy density, cycle life, cost, and safety and environmental protection. In practical engineering applications, a single energy storage technology (such as lithium batteries) is often used for dispatching. However, a single technology cannot simultaneously meet the multi-timescale requirements of millisecond-level power response, hourly energy time shift, and cross-seasonal energy storage. To resolve this contradiction, Hybrid Energy Storage System (HESS) has been proposed, which combines two or more energy storage technologies.

[0003] Existing hybrid energy storage scheduling methods typically allocate total power demand to different energy storage devices according to pre-set fixed rules, or use a centralized optimization model to uniformly calculate the charging and discharging power of each energy storage device. Some methods use low-pass filtering or model predictive control to divide power demand into time scales, while others receive constraints and target weights input by the scheduler through a parameter configuration interface.

[0004] However, fixed rules are difficult to adapt to dynamic changes in electricity prices, load, renewable energy output, and energy storage status; centralized optimization methods are usually based on the optimization objectives of a single entity, making it difficult to reflect the interests of different energy storage entities; traditional parameter configuration methods also struggle to directly understand natural language objectives such as "prioritizing battery life protection" or "activating power supply protection mode." Furthermore, existing frequency band allocation methods are often independent of multi-entity economic decision-making, potentially leading to a mismatch between economic optimization results and the response speed of energy storage devices. Summary of the Invention

[0005] The main purpose of this application is to provide a hybrid energy storage scheduling method and system based on LLM agents and game theory, which aims to solve the technical problems of existing technologies lacking in-depth differential analysis of the characteristics of hybrid energy storage systems and ignoring the game relationship between various devices in hybrid energy storage systems, resulting in poor autonomous scheduling capability and uneven resource allocation of hybrid heterogeneous energy storage systems.

[0006] To achieve the above objectives, this application provides a hybrid energy storage scheduling method based on LLM agents and game theory, the method comprising the following steps: The collected multi-source heterogeneous data is preprocessed, and a total power command is generated based on the preprocessed multi-source heterogeneous data. In response to a natural language scheduling instruction sent by a user, a scheduling parameter set is generated based on the natural language scheduling instruction, the scheduling parameter set including quantization constraints and target weights; Based on the total power command and the scheduling parameter set, the basic charge and discharge power curve in the converged state is obtained by performing master-slave game iteration through the leader agent and multiple follower agents. The leader agent and multiple follower agents all use a large language model as the inference engine. The multiple follower agents correspond to different types of energy storage devices. The follower agents include capacity followers, power followers and ultra-long time followers. The total power command is decomposed into multiple frequency band power components, and a target scheduling command sequence is generated based on the multiple frequency band power components and the basic charge and discharge power curve. The target scheduling instruction sequence is sent to the controller of each energy storage device, so that each controller controls the corresponding energy storage device to perform charging and discharging operations according to the target scheduling instruction sequence.

[0007] Furthermore, to achieve the above objectives, this application also proposes a hybrid energy storage scheduling system based on LLM agents and game theory, which applies the hybrid energy storage scheduling method based on LLM agents and game theory described above. The hybrid energy storage scheduling system based on LLM agents and game theory includes: The data processing module is used to preprocess the collected multi-source heterogeneous data and generate a total power command based on the preprocessed multi-source heterogeneous data. The instruction response module is used to respond to the natural language scheduling instruction sent by the user and generate a scheduling parameter set based on the natural language scheduling instruction. The scheduling parameter set includes quantization constraints and target weights. The master-slave game module is used to obtain the basic charge and discharge power curve in the converged state by performing master-slave game iteration through a leader agent and multiple follower agents based on the total power command and the scheduling parameter set. The leader agent and the multiple follower agents all use a large language model as the inference engine. The multiple follower agents correspond to different types of energy storage devices. The follower agents include capacity followers, power followers and ultra-long time followers. The frequency band decomposition module is used to decompose the total power command into multiple frequency band power components, and generate a target scheduling command sequence based on the multiple frequency band power components and the basic charge and discharge power curve. The energy storage scheduling module is used to send the target scheduling instruction sequence to the controller of each energy storage device, so that each controller controls the corresponding energy storage device to perform charging and discharging operations according to the target scheduling instruction sequence.

[0008] Furthermore, to achieve the above objectives, this application also proposes a hybrid energy storage scheduling device based on LLM agents and game theory. The device includes: a memory, a processor, and a hybrid energy storage scheduling program based on LLM agents and game theory stored in the memory. The processor is used to run the hybrid energy storage scheduling program based on LLM agents and game theory. The computer program is configured to implement the steps of the hybrid energy storage scheduling method based on LLM agents and game theory as described above.

[0009] This application reduces the impact of outliers and timescale differences on scheduling results by preprocessing multi-source heterogeneous data; it enables unstructured objectives such as lifetime protection, power supply priority, and profit preference to enter the subsequent optimization process by converting natural language scheduling instructions into quantified constraints and target weights; it incorporates electricity prices and leasing strategies, energy storage device operating status, and response strategies of each entity into an iterative closed loop through a leader agent and multiple follower agents to form a basic charge-discharge power curve that takes into account both entity profits and device constraints; it further matches power-type, capacity-type, and ultra-long-duration energy storage devices according to the timescale of power changes through frequency band decomposition, so that the economic scheduling baseline obtained by the game is subject to secondary correction by the physical response capabilities of heterogeneous energy storage; and it issues target scheduling instructions to each controller. This invention enables autonomous scheduling of hybrid heterogeneous energy storage systems by using a sequence of commands. This avoids the limitations of focusing solely on economic optimization while ignoring physical response characteristics or focusing solely on frequency allocation while ignoring the interests of stakeholders. It forms a collaborative scheduling mechanism that takes into account scheduling intentions, multi-stakeholder coordination, and the executability of heterogeneous energy storage. Because this invention significantly reduces the overall levelized cost and comprehensive electricity cost of the energy storage system through heterogeneous energy storage task decomposition and master-slave game equilibrium, it achieves a balance of interests between energy storage operators and various energy storage devices through leader-follower master-slave game, effectively improving the benefits of all parties and the overall system efficiency. By allocating power commands according to the time scale through wavelet packet decomposition, it fully leverages the technical advantages of power-type, capacity-type, and ultra-long-duration energy storage, significantly improving the annual average utilization rate of the energy storage system and effectively suppressing net load fluctuations. Attached Figure Description

[0010] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0011] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0012] Figure 1 This is a flowchart illustrating the first embodiment of the hybrid energy storage scheduling method based on LLM agents and game theory in this application. Figure 2 This is a flowchart illustrating the second embodiment of the hybrid energy storage scheduling method based on LLM agents and game theory in this application. Figure 3 This is a flowchart illustrating the third embodiment of the hybrid energy storage scheduling method based on LLM agents and game theory in this application. Figure 4 This is a schematic diagram of the interaction timing of each agent in one embodiment of the hybrid energy storage scheduling method based on LLM agents and game theory in this application; Figure 5 This is a schematic diagram of the module structure of the hybrid energy storage scheduling system based on LLM agents and game theory in an embodiment of this application; Figure 6 This is a schematic diagram of the device structure of the hardware operating environment involved in the hybrid energy storage scheduling method based on LLM agents and game theory in the embodiments of this application.

[0013] The purpose, features, and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0014] It should be understood that the specific embodiments described herein are merely illustrative of the technical solutions of this application and are not intended to limit this application.

[0015] To better understand the technical solution of this application, a detailed description will be provided below in conjunction with the accompanying drawings and specific implementation methods.

[0016] It should be noted that the executing entity in this application embodiment can be a computing service device with data processing, network communication, and program execution functions, such as a server, computer, edge computing device, energy storage management device, or a computing cluster composed of multiple computing devices. This executing entity can interact with meteorological data platforms, power trading platforms, power grid dispatch centers, and the controllers of various energy storage devices. The following description uses a hybrid energy storage dispatching device based on LLM agents and game theory (hereinafter referred to as the dispatching device) as an example to illustrate this embodiment and the following embodiments.

[0017] Based on this, the first embodiment of this application provides a hybrid energy storage scheduling method based on LLM agents and game theory, referring to... Figure 1 , Figure 1 This is a flowchart illustrating the first embodiment of the hybrid energy storage scheduling method based on LLM agents and game theory in this application.

[0018] In this embodiment, the hybrid energy storage scheduling method based on LLM agents and game theory includes: Step S10: Preprocess the collected multi-source heterogeneous data and generate a total power command based on the preprocessed multi-source heterogeneous data.

[0019] It should be noted that multi-source heterogeneous data refers to power system operation data originating from different monitoring objects and possessing different sampling frequencies and data formats, such as meteorological data, electricity price signals, load forecast curves, carbon quota trading prices, and operational status data of various energy storage devices. Preprocessing refers to the cleaning, alignment, and normalization operations performed on the raw data. Total power command refers to the net power value that the energy storage system needs to respond to within the current dispatch cycle.

[0020] In some embodiments, the dispatching equipment collects raw multi-source heterogeneous data from the target power grid and the hybrid energy storage system; identifies and removes outliers from the raw multi-source heterogeneous data based on preset anomaly judgment rules to obtain outlier-removed data; performs linear interpolation on the missing data in the outlier-removed data to obtain complete multi-source heterogeneous data; normalizes the complete multi-source heterogeneous data according to the data range corresponding to each type of data to obtain preprocessed multi-source heterogeneous data; and performs net power calculation on the preprocessed multi-source heterogeneous data to obtain the total power command.

[0021] In its implementation, the dispatching equipment collects multi-source data every 1 to 15 minutes: meteorological data (irradiance, wind speed, ambient temperature, and 72-hour weather forecast), electricity price signals (nodal marginal price, time-of-use price table, and real-time spot price), ultra-short-term and short-term load forecast curves obtained from the power grid dispatching center, daily carbon quota trading prices, and the operating status of various energy storage devices (SOC, SOH, maximum charge / discharge power, temperature, and cycle count for capacity-type energy storage; speed, available power, and bearing temperature for power-type energy storage; tank pressure, remaining energy, and efficiency coefficient for ultra-long-term energy storage; and thermal temperature, thermal power, and thermoelectric conversion efficiency for thermal energy storage). The collected data undergoes outlier removal, missing value imputation (linear interpolation), and normalization before being stored in a time-series database and a vector database for use by the decision-making agent.

[0022] Based on the collected power grid load Photovoltaic power generation and wind power The system calculates the total power command: This instruction represents the net power value required for the overall response of the energy storage system (positive values ​​indicate discharge, negative values ​​indicate charging), and it serves as the input signal for subsequent high-frequency smoothing of power-type energy storage and wavelet packet decomposition in step six.

[0023] Step S20: In response to the natural language scheduling instruction sent by the user, generate a scheduling parameter set based on the natural language scheduling instruction, the scheduling parameter set including quantization constraints and target weights.

[0024] It should be noted that natural language scheduling instructions refer to operational goals or constraints expressed by users in natural language, such as "prioritize extending battery life," "activate typhoon power supply mode," or "reduce power purchase during peak hours." Quantitative constraints refer to the numerical range or logical conditions that can be directly handled by the optimization model. Objective weights refer to the relative importance of different optimization objectives in the comprehensive decision-making process.

[0025] In practice, the scheduling device inputs natural language scheduling instructions into the policy commander agent, which uses a large language model as its inference engine.

[0026] Users can input natural language scheduling objectives through the human-computer interaction interface, such as: "There will be typhoon weather in the next three days. Please activate the power supply mode to prioritize the important loads in the industrial park, while minimizing the degradation of lithium batteries."

[0027] The strategy commander agent autonomously executes the following sub-tasks using the ReAct paradigm: Consideration: Users request a power-saving mode, typhoon warnings, and reduced lithium battery degradation. We need to retrieve the historical definition of "power-saving mode" from the memory; if it's not found, we should generate new parameters.

[0028] Action: Call `memory_retrieval_tool` (RAG-based vector retrieval) to search for historical "power preservation mode" configurations. If not found, call `constraint_generator_tool` to convert "reduce decay" into quantized constraints (maximum discharge depth limited to 40%, SOC lower limit increased to 60%). Simultaneously, call `weight_allocator_tool` to convert the ambiguous target into a weight vector (revenue weight 0.3, lifetime weight 0.7).

[0029] Observation: Parameter set successfully generated: power preservation mode activated, SOC lower limit = 60%, discharge depth limit = 40%, weight (0.3, 0.7).

[0030] Output: Broadcast the above parameters to all follower agents in natural language and reply to the user on the interface: "Your command has been parsed. Power preservation mode will be activated. Lithium battery life protection has been activated."

[0031] Step S30: Based on the total power command and the scheduling parameter set, perform master-slave game iteration through the leader agent and multiple follower agents to obtain the basic charge and discharge power curve in the converged state.

[0032] It should be noted that the basic charge / discharge power curve refers to the initial power plan formed by each energy storage device within the target scheduling period after the master-slave game reaches a convergence state. The master-slave game refers to the Stackelberg game.

[0033] Both the leader agent and the multiple follower agents use a large language model as the inference engine. The multiple follower agents correspond to different types of energy storage devices, including capacity-type followers, power-type followers, and ultra-long-duration followers.

[0034] A leader agent is a decision-making agent that preemptively formulates electricity pricing and energy storage capacity leasing strategies in a master-slave game. A follower agent is a decision-making agent that determines response capacity and charging / discharging power based on the leader's strategy and the operating status of the corresponding energy storage device.

[0035] Among them, capacity-type followers can correspond to lithium batteries or flow batteries, power-type followers can correspond to flywheels or supercapacitors, and ultra-long-duration followers can correspond to compressed air energy storage or hydrogen energy storage.

[0036] Understandably, master-slave game theory can incorporate the leader's pricing and resource needs, as well as the followers' device status, operating costs, and willingness to respond, into alternating iterations, so that the basic charge and discharge power curve reflects the strategic relationships between different entities.

[0037] The Stackelberg game constructed in this embodiment is essentially a two-level optimization model, comprising an upper-level leader and lower-level followers. The upper-level leader (energy storage operator) guides the charging and discharging behavior of various types of followers by setting electricity prices and leasing rates to maximize its own profits. Each follower (capacity-type, power-type, ultra-long-duration, and thermal storage) independently decides its charging and discharging power based on its own profit maximization under a given electricity price signal. The upper-level decisions directly affect the lower-level profit function, and the collective response of the lower-level leaders, in turn, determines the actual profit of the upper-level leaders. This process iterates until a Stackelberg equilibrium is reached.

[0038] Step S40: Decompose the total power command into multiple frequency band power components, and generate a target scheduling command sequence based on the multiple frequency band power components and the basic charge and discharge power curve.

[0039] It should be noted that the frequency band power component refers to the power variation components corresponding to the total power command in different frequency ranges. The target scheduling command sequence refers to the charging and discharging control targets organized according to the energy storage device identification and scheduling period, which may include the target power, charging and discharging direction, and execution period.

[0040] It is understood that this embodiment can distinguish between rapid fluctuations, medium-time scale changes and long-period trends in total power commands through frequency band decomposition, enabling energy storage devices with different dynamic characteristics to undertake matching power tasks.

[0041] In its implementation, the scheduling equipment employs wavelet packet decomposition, filter bank decomposition, or other signal processing methods capable of obtaining power components across multiple frequency ranges to perform frequency band decomposition on the total power command. Based on the center frequency or characteristic period of each frequency band power component, the equipment allocates rapidly changing power components to power-type followers, power components on a minute-to-hour scale to capacity-type followers, and long-period power components to ultra-long-duration-time-type followers. Each follower superimposes its allocated power component onto the corresponding baseline charge / discharge power curve generated during the game phase, generating a sub-scheduling command sequence corresponding to the energy storage type. The equipment then arranges the sub-scheduling command sequences according to the energy storage device identifier and scheduling time period to form the target scheduling command sequence.

[0042] Step S50: Send the target scheduling instruction sequence to the controller of each energy storage device, so that each controller controls the corresponding energy storage device to perform charging and discharging operations according to the target scheduling instruction sequence.

[0043] It should be noted that an energy storage device controller refers to a control unit that can receive power commands and control the energy storage device to perform charging, discharging or standby operations. It can be an energy storage converter controller, an energy management controller or a local controller for energy storage devices.

[0044] In practice, the dispatching equipment splits the target dispatching command sequence according to the energy storage device identifier, encodes each sub-sequence into a data format recognizable by the corresponding controller, and sends it to the corresponding controller via wired or wireless communication network. Before execution, the controller checks whether the target power exceeds the rated power, whether the state of charge is within the allowable range, and whether the communication data is complete. After the checks are passed, the controller adjusts the energy storage converter, motor, compressor, or expander according to the target power and execution period, enabling the corresponding energy storage device to perform charging and discharging operations. The controller can also feed back the actual power, energy storage status, and equipment temperature to the hybrid energy storage dispatching equipment for use in the next dispatching cycle.

[0045] This embodiment converts user objectives into computable constraints and weights through natural language parsing, forms a basic power plan that takes into account the strategic relationships of multiple parties through master-slave game theory, matches power tasks with the response time scales of different energy storage devices through frequency band decomposition, and completes the physical execution of scheduling results through the controller interface. Therefore, this embodiment significantly reduces the overall levelized cost and comprehensive electricity cost of the energy storage system through heterogeneous data decomposition and master-slave game equilibrium, and improves the adaptability and synergy of hybrid energy storage scheduling through master-slave game theory between leaders and followers.

[0046] Based on the first embodiment described above, in the second embodiment of this application, the same or similar content as the first embodiment can be referred to the above description, and will not be repeated hereafter.

[0047] Based on this, please refer to Figure 2 , Figure 2 This is a flowchart illustrating the second embodiment of the hybrid energy storage scheduling method based on LLM agents and game theory in this application. In this embodiment, step S30 further includes: Step S301: Invoke the leader agent to generate the electricity price and leasing strategy message for the current round based on the total power command and the scheduling parameter set.

[0048] It should be noted that electricity price and leasing strategy messages refer to strategy information sent by the leader agent to the follower agents, which may include time-of-use electricity price sequences, leasing rates, expected leasing capacity, and applicable time periods of the strategy.

[0049] In the specific implementation, the scheduling device calls the leader agent to read the total power command, quantization constraints, and target weights. Based on the current predicted electricity price, energy storage resource gap, and the follower response in the previous round, the leader agent calls the electricity price prediction tool and optimization solution tool to form a time-of-use electricity price sequence and energy storage capacity leasing scheme covering the target scheduling cycle, and encapsulates the above information into the electricity price and leasing strategy message for the current round.

[0050] Step S302: Broadcast the electricity price and leasing strategy message of the current round to each follower agent participating in the master-slave game.

[0051] It should be understood that sending policy messages with the same round identifier to all followers can prevent followers from responding based on different round policies.

[0052] In the specific implementation, the scheduling device adds a game round identifier, a target scheduling cycle identifier, and a sending timestamp to the electricity price and leasing strategy messages, and then sends them to capacity-type followers, power-type followers, and ultra-long-duration followers via a message bus or intelligent agent communication interface. Each follower verifies the round identifier and message integrity, and uses the verified message as the decision input for the current round.

[0053] Step S303: Invoke each follower agent to make a charging and discharging decision for the current round based on the current round's electricity price and leasing strategy message, the operating status of the corresponding energy storage device, and the scheduling parameter set, and generate a follower response message based on the charging and discharging decision result.

[0054] It should be noted that the follower response message refers to the structured decision result returned by the follower agent after solving the current round's strategy, including the response capacity and the response charge / discharge power curve.

[0055] In the specific implementation, the scheduling device calls upon each follower agent to parse the current round's policy message and reads the available capacity, current energy storage state, maximum charge / discharge power, energy conversion efficiency, and temperature of the corresponding energy storage device from the preprocessed multi-source heterogeneous data. Each follower selects a local optimizer based on the energy storage type, calculates candidate charge / discharge strategies by combining quantization constraints and target weights, filters out candidate strategies that violate operational constraints, and then encapsulates the capacity and power curves corresponding to the selected strategy into a follower response message.

[0056] Step S304: Send the follower response message to the leader agent, call the leader agent to calculate the leader's revenue for the current round based on the follower response message, and update the electricity price and leasing strategy messages according to the leader's revenue for the current round.

[0057] In its implementation, the scheduling device sends response messages from each follower to the leader agent. The leader agent aggregates the response capacity and charge / discharge power curves, calculates the deviation between the aggregated response and the total power command, and calculates the current round's cumulative revenue based on time-of-use electricity revenue, leader energy storage operating costs, and leasing costs. The leader agent then re-invokes the optimization tool based on changes in cumulative revenue and power deviation, updates the electricity price and leasing rate, and generates the next round's strategy message.

[0058] Step S305: If the preset convergence condition is not met, proceed to the next round of master-slave game iteration based on the updated electricity price and leasing strategy messages; Step S306: If the preset convergence condition is met, obtain the basic charge and discharge power curve based on the follower response message of the current round.

[0059] It should be noted that the preset convergence condition refers to the pre-set game termination criterion, which may include the leader's payoff rate not exceeding the payoff rate threshold, the follower's power curve rate not exceeding the strategy change threshold, and the above conditions being met for multiple consecutive rounds.

[0060] In the specific implementation, the scheduling device calculates the rate of change of the leader's payout and the rate of change of the follower's power curve between the current round and the previous round. When either rate of change fails to meet the corresponding threshold, the device uses the updated strategy message to enter the next iteration. When all rates of change continuously reach a preset number of rounds, the device terminates the game and extracts the response charging and discharging power curves of each energy storage device from the follower response messages of the current round to form the basic charging and discharging power curves. For example, the payout change threshold can be set based on the allowable economic evaluation fluctuations, and the strategy change threshold can be set based on the power control accuracy of the energy storage device.

[0061] Furthermore, the leader agent is configured to perform the following steps during the master-slave game iteration process: Step S3011: Extract historical electricity price sequences from the preprocessed multi-source heterogeneous data, and input the historical electricity price sequences into the pre-trained electricity price prediction model to obtain the predicted electricity price sequence within the target scheduling period.

[0062] It should be noted that the electricity price prediction model can employ recurrent neural networks, temporal convolutional networks, or attention-based temporal prediction networks. The model uses historical electricity prices and concurrent influence features as training samples, with the actual electricity price for the corresponding future period as the label, and the error between the predicted and actual electricity prices as the training loss. After training on the training set and selecting a validation set, a pre-trained model is obtained. During inference, the model outputs the predicted electricity price for each period within the target scheduling cycle.

[0063] Step S3012: Construct a leader optimization model based on the predicted electricity price sequence, the total power command, and the scheduling parameter set.

[0064] It should be noted that the leader optimization model refers to a higher-level optimization model that uses electricity pricing strategy and energy storage capacity leasing strategy as decision variables and leader revenue as the optimization objective. The optimization objective of the leader optimization model is to maximize leader revenue, and the constraints of the leader optimization model include the state of charge constraint and charging / discharging power constraint of the leader's energy storage device.

[0065] Step S3013: Using the historical response messages of the follower agent as the initial optimization parameters, call the mixed integer linear programming solver to solve the leader optimization model and obtain the solution status identifier, candidate electricity price strategy, candidate energy storage capacity leasing strategy and corresponding leader revenue; Step S3014: If the solution status indicator indicates that the solution is successful, select the target electricity price strategy and the target energy storage capacity leasing strategy from the candidate electricity price strategy and the candidate energy storage capacity leasing strategy based on the leader's reward, generate the electricity price and leasing strategy message, and broadcast the electricity price and leasing strategy message to each follower agent participating in the master-slave game.

[0066] In practical implementation, during the master-slave game, the upper-level optimization problem in the two-layer optimization model based on the master-slave game is optimized by the leader agent. The leader agent can be configured (representing the energy storage operator) to own a large centralized energy storage power station (e.g., a 1MW / 2MWh lithium battery) and the right to lease capacity from followers (distributed energy storage). Its goal is to maximize its own revenue while satisfying the power supply guarantee mode constraint. The revenue function is as follows: in, Represents the cumulative gain obtained by the leader agent throughout the entire scheduling cycle, with the subscript... Represents the leader intelligent agent; This represents the total number of scheduling periods, T=96 (one period every 15 minutes for the next 24 hours); Indicates the duration of a single scheduling period; The electricity price set by the leader for the followers (decision variable); Indicates the sequence number of the scheduling period; This represents a set of follower agents, including capacity-type followers (lithium batteries), power-type followers (flywheels), and ultra-long-duration followers (compressed air), etc. Let be the discharge power of the i-th follower during the t-th scheduling period; This represents the unit discharge operating cost of the leader agent's own energy storage device. =0.05 yuan / kWh; For rental rates (decision variables); The capacity leased to followers (a decision variable determined by follower responses); This represents the discharge amount of the leader agent's own energy storage device during the t-th scheduling period.

[0067] The leader's ReAct decision-making process is as follows: Consideration: "The current peak-valley electricity price difference is 0.9 yuan / kWh (1.2 yuan during peak hours and 0.3 yuan during off-peak hours), but the power supply guarantee mode requires a SOC of no less than 60%, reducing the available capacity of its own energy storage. A pricing scheme for charging followers needs to be developed. The peak-hour price should be lower than the grid peak price to attract followers to discharge, but it cannot be too low; the off-peak price should be higher than the grid off-peak price to encourage followers to charge. A leasing rate of 0.05 yuan / kWh is reasonable."

[0068] action: 1) Call price_forecast_tool, input parameters: forecast duration 72 hours, output granularity 15 minutes. This tool internally uses the Transformer model and returns the time-of-use electricity price forecast sequence p_forecast[t] for the next 72 hours.

[0069] 2) Call cost_solver (a mixed-integer linear programming solver based on Gurobi), with the following inputs: the leader agent's revenue function, constraints (the leader's own energy storage SOC constraint and power limit), and the follower's historical response model (set as the initial guess). Solve to obtain the initial pricing scheme p_L_initial[t] and the rental rate r_initial=0.05.

[0070] 3) Solver output: Peak electricity price 0.95 yuan / kWh, off-peak electricity price 0.25 yuan / kWh, and normal electricity price 0.60 yuan / kWh.

[0071] Observation: "Solution successful, pricing scheme generated".

[0072] Output: Broadcast a natural language message to all follower agents, such as: "I propose the following electricity pricing scheme for the next 24 hours: peak hours (10:00-12:00, 17:00-19:00) electricity price 0.95 yuan / kWh, off-peak hours (23:00-07:00) electricity price 0.25 yuan / kWh, and normal hours electricity price 0.60 yuan / kWh. The rental rate is 0.05 yuan / kWh. Please provide your response capacity and power curves within 15 minutes."

[0073] Furthermore, the capacity-type follower is configured to perform the following steps during the master-slave game: Step S3031: Parse the electricity price and leasing strategy message of the current round to obtain the target electricity price strategy and target energy storage capacity leasing strategy of the current round, and extract the energy storage type identifier and operating status of the corresponding energy storage device from the preprocessed multi-source heterogeneous data.

[0074] It should be noted that the energy storage type identifier refers to a data identifier used to characterize the technology type of an energy storage device. Capacity-type energy storage devices can be lithium-ion battery energy storage devices or flow battery energy storage devices. The operating status of a capacity-type energy storage device can include state of charge, state of health, rated capacity, available capacity, maximum charge / discharge power, charge / discharge efficiency, temperature, and cycle count.

[0075] Step S3032: Construct a follower optimization model based on the target electricity price strategy, target energy storage capacity leasing strategy, energy storage type identifier, and operating status.

[0076] It should be noted that the optimization objective of the follower optimization model is to maximize the follower's profit, and the constraints of the follower optimization model include charging and discharging power constraints, available capacity constraints of the energy storage device, energy storage state constraints, and energy storage state transition constraints.

[0077] It should be noted that energy storage state constraints can be upper and lower limits of state of charge constraints; energy storage state transition constraints refer to the constraint relationship between adjacent scheduling periods where the state of charge changes according to charging power, discharging power, charging and discharging efficiency and rated capacity.

[0078] Step S3033: Using the historical response messages of the corresponding follower agent as the initial optimization parameters, call the local optimizer that matches the energy storage type of the corresponding energy storage device to solve the follower optimization model, and obtain the solution status identifier, candidate charging and discharging strategies, candidate response capacity, and follower benefits corresponding to each candidate charging and discharging strategy.

[0079] It should be noted that a local optimizer is a computational tool used to solve the charging and discharging strategy of a single follower. It can be a mathematical programming solver, a dynamic programming solver, or a deep reinforcement learning network trained offline.

[0080] Step S3034: If the solution status indicator indicates that the solution is successful, select a target charging and discharging strategy from the candidate charging and discharging strategies based on the follower's reward, generate a follower response message according to the target charging and discharging strategy and the corresponding candidate response capacity, and send the follower response message to the leader agent.

[0081] In the specific implementation, after receiving the leader's price signal, each follower agent solves its own profit maximization problem to determine the optimal charging and discharging power. The objective function is: in, This represents the charging power sequence of the energy storage device corresponding to the i-th follower agent during the scheduling period. and discharge power sequence Let these be decision variables, and under the constraints of the follower optimization model, maximize the cumulative return. Reaching the maximum; Indicates the first The cumulative gain of a follower agent within a scheduling cycle; Indicates the first The energy storage device corresponding to each follower agent is in the first... Charging power during each scheduling period; Indicates the first The unit lifetime degradation cost coefficient of the energy storage device corresponding to each follower agent. Indicates the first The energy storage device corresponding to each follower agent is in the first... The charging and discharging power during each scheduling period.

[0082] The constraints on follower agents include power constraints, SOC constraints, SOC dynamic equations, and lifetime decay constraints (see the response processes of various follower types below for details).

[0083] Capacity follower represents a user-side lithium battery system (parameters: capacity 500kW / 1000kWh, maximum charge / discharge power 250kW, charge / discharge efficiency 90%, degradation cost factor). =0.01 yuan / kWh). After receiving the leader's electricity price, execute the ReAct loop: The thought process is as follows: "The off-peak electricity price of 0.25 yuan / kWh is lower than my charging cost (grid off-peak price 0.3 yuan / kWh × reciprocal of efficiency ≈ 0.33 yuan / kWh? Actual calculation: charging cost = electricity purchase price / efficiency = 0.25 / 0.9 ≈ 0.278 yuan / kWh, while my implicit charging cost baseline is 0.23 yuan / kWh, therefore it is advantageous. The current SOC is 55%, and the power reserve mode requires 60%, so charging is necessary first. The depth of discharge should be controlled within 40%. I need to solve the charge and discharge sequence for the next 24 hours." The internal reasoning also includes numerical calculations for cost comparison.

[0084] Action: Call the three tools in sequence: 1) `soc_prediction_tool`: Input the current SOC = 0.55, candidate charge / discharge sequences (100 random sequences generated by a heuristic algorithm), and output the SOC trajectory corresponding to each sequence. This tool internally uses a simple integral model. in, This indicates the state of charge of a capacity-type energy storage device at the beginning of the t-th scheduling period. Indicates the charging efficiency of a capacity-type energy storage device; This represents the charging power of the capacity-type energy storage device during the t-th scheduling period; This represents the discharge power of the capacity-type energy storage device during the t-th dispatch period; Indicates the discharge efficiency of a capacity-type energy storage device; This indicates the rated energy storage capacity of a capacity-type energy storage device.

[0085] 2) degradation_model: Input power sequence, output lifetime decay cost. The model uses a linear cumulative model: in, This represents the lifespan degradation cost incurred by an energy storage device during a single dispatch period. =0.01 yuan / kWh; This represents the unit lifespan degradation cost coefficient of an energy storage device; This indicates the charging and discharging power of the energy storage device during the current dispatch period; 3) Local Optimizer: Based on a Deep Q-Network (DQN), the state space is (current SOC, current electricity price, number of remaining time slots), and the action space is a discrete power level (-250kW to 250kW, step size 10kW, 51 actions in total). An offline-trained DQN model (3 fully connected layers, 128 hidden layers, learning rate 0.001, experience replay capacity 10000) is used. The current state is input, and the optimal action sequence is output. The solution process is completed within 0.2 seconds.

[0086] Observation: The optimal strategy is as follows: charging period 00:00-03:00, power 120kW; discharging period 11:00-13:00, power 80kW; discharging period 17:00-19:00, power 60kW. The estimated total discharge is 240kWh, the charging amount is 360kWh, and the net profit is 3200 yuan.

[0087] Output: Natural Language Response Leader: "I accept your electricity pricing plan. My response plan: Charging period 00:00-03:00, power 120kW; Discharging period 11:00-13:00, power 80kW; Discharging period 17:00-19:00, power 60kW. Estimated total discharge 240kWh, charging 360kWh. Requesting confirmation."

[0088] It should be noted that power-type followers do not participate in electricity price negotiations, but instead run a separate thread continuously. Their data flow is as follows: the perception layer collects the power of the grid-connected points 100 times per second. ,calculate .

[0089] in, This indicates the load power at the current sampling time; This represents the photovoltaic power output at the current sampling time; This indicates the wind power output at the current sampling time; This represents the net load power obtained after deducting the output power of photovoltaic and wind power from the load power, and is used as the total power command. All power values ​​use the same sampling time and unit of measurement. A value greater than 0 indicates that the load demand exceeds the combined output of solar and wind power. <0 indicates that the combined output of photovoltaic and wind power exceeds the load demand.

[0090] Then, wavelet packet decomposition is performed every 10ms. The decomposition uses Daubechies-4 wavelets, with 4 levels of decomposition, to extract the highest frequency subband coefficients. The negative value of this high-frequency subband coefficient is then used as the flywheel's charging and discharging command (flywheel response range -100kW~100kW). The flywheel's internal model adjusts the motor torque according to the command, with an actual response delay of <5ms. Every second, the flywheel reports its status to the strategy commander: "Flywheel power = xxkW, rotor speed = xxrpm, temperature = xx°C".

[0091] Furthermore, the ultra-long-duration follower is configured to perform the following steps during the master-slave game: Step S3131: Extract the target electricity price strategy and target energy storage capacity leasing strategy for the current round from the current round's electricity price and leasing strategy message, and obtain the energy storage pressure range, charging and discharging power range, and energy conversion efficiency of the ultra-long-duration energy storage device.

[0092] It should be noted that ultra-long-duration energy storage devices refer to energy storage devices capable of storing and releasing energy over periods of several hours, days, or longer, and can be compressed air energy storage devices. The energy storage pressure range can be from the minimum to the maximum operating pressure allowed by the compressed air facility.

[0093] Step S3132: Extract meteorological forecast data and load forecast data for the preset planning period from the preprocessed multi-source heterogeneous data.

[0094] In practice, the scheduling equipment reads meteorological forecast data such as wind speed, irradiance, ambient temperature, and weather type for the preset planning period from the pre-processed multi-source heterogeneous data, and also reads the load forecast curve for the same period. The equipment estimates the wind power and photovoltaic output trends based on the meteorological forecast data, and then provides the new energy output trends and load forecast curves to ultra-long-duration followers as the external state for long-term decision-making.

[0095] Step S3133: Discretize the energy storage pressure range and the charge / discharge power range to obtain multiple pressure states and multiple charge / discharge power levels.

[0096] It should be noted that pressure state refers to the discrete state formed after dividing the continuous energy storage pressure range. Charge / discharge power level refers to the multiple selectable decision values ​​formed after dividing the continuous power range.

[0097] Step S3134: Construct a state transition relationship and a long-term scheduling constraint set based on the multiple pressure states, the multiple charging and discharging power levels, the energy conversion efficiency, and the scheduling parameter set.

[0098] It should be noted that the state transition relationship refers to the correspondence between the current pressure state and the next pressure state of an ultra-long-duration energy storage device after selecting a charge / discharge power level. The ultra-long-duration scheduling constraint set may include pressure upper and lower limits, charge / discharge power range, energy conservation, conversion efficiency, continuous operating time, and capacity leasing constraints.

[0099] Step S3135: Based on the state transition relationship and the ultra-long-term scheduling constraint set, cumulatively solve the charging and discharging revenue corresponding to the target electricity price strategy and the target energy storage capacity leasing strategy to obtain the target long-term charging and discharging power sequence; Step S3136: Determine the response capacity and response charge / discharge power curve of the ultra-long-duration energy storage device based on the target long-cycle charge / discharge power sequence, generate a follower response message, and send the follower response message to the leader agent.

[0100] In its implementation, the ultra-long-duration follower receives the rental rate (0.05 yuan / kWh) and the leader's desired rental capacity (e.g., the leader requests an additional 100kWh for peak hours the following day). The ultra-long-duration follower executes: Thinking: "The 7-day weather forecast shows high wind speeds at night, which is suitable for charging; high electricity prices during the daytime peak hours, which is suitable for discharging. The rental rate of 0.05 yuan / kWh is low, but there is still profit when combined with arbitrage for self-use. I need to develop a 7-day charging and discharging plan."

[0101] Action: Invoke the seasonal_planner tool, which is based on dynamic programming (DP). The state variable is the gas tank pressure (0-10MPa, discretized into 100 states), and the decision variable is the charging and discharging power (-500kW~500kW, discretized into 20 levels). The objective function is to maximize the cumulative revenue over 7 days (including rental income and peak-valley arbitrage). The time complexity of DP is O(7×24×100×20)≈336,000, which can be completed within 1 second.

[0102] Observation: The optimal plan is to charge at 300kW from 02:00 to 06:00 daily and discharge at 250kW from 17:00 to 21:00. Of this, 200kWh of discharge capacity can be provided daily, with 100kWh leased to the leader and 100kWh for self-use.

[0103] Output: Reply to the leader agent: "I agree to the leasing plan. I will provide 100kWh of discharge capacity per day, from 17:00 to 21:00, at a power of 25kW (because leased power = capacity / 4 hours = 25kW). The remaining 100kWh will be used by myself."

[0104] In one embodiment, after the scheduling device invokes the policy commander agent to collect the responses of all followers, it enters an iterative loop. Let the leader's reward after the k-th iteration be... Each follower's strategy is as follows: .calculate: in, Indicates the first Wheel relative to the first The relative rate of change in the leader's earnings in the cycle; and They represent the first Wheel and the first The cumulative reward of the leader within the same objective scheduling cycle, calculated in round-robin mode; subscript This represents the leader intelligent agent.

[0105] This represents the relative change rate of the charging and discharging power strategy of the i-th follower agent in round k compared to round k-1; and Let each represent a charge / discharge power sequence generated by the i-th follower agent in the corresponding round, covering the same target scheduling period. Elements at the same position in each sequence correspond to the same scheduling time period. This represents the L2 norm, which is the square root of the sum of the squares of the elements in the sequence.

[0106] like ≤0.01 and all If the value is ≤0.02 and satisfies this condition for 3 consecutive rounds, then the convergence is determined.

[0107] Example of actual iteration process: Round 1: Leader pricing peak 0.95 / trough 0.25, follower response total 450kWh; leader's revenue 12,000 yuan.

[0108] Round 2: The leader adjusts the pricing based on the response, with a peak of 0.93 and a trough of 0.27. The total follower response volume is 480 kWh. The leader's revenue is 12,500 yuan. The relative change rate of the leader's revenue in Round 2 is as follows. =4.2%>1%, not converged.

[0109] Round 3: Leader pricing peak 0.92 / trough 0.28, follower response total 500kWh; leader revenue 12,700 yuan, relative change rate of leader revenue in Round 3. =1.6%>1%.

[0110] Round 4: Leader pricing peak 0.91 / trough 0.29, follower response total 510kWh; leader revenue 12,800 yuan, relative change rate of leader revenue in Round 4. =0.8%≤1%, and the change in follower strategy is <2%, which satisfies the condition, but requires 3 consecutive rounds and will not converge at this time.

[0111] Round 5: The leader's pricing remained at a peak of 0.91 / trough of 0.29 (based on the previous successful strategy), with a total follower response of 512 kWh; the leader's revenue was 12,830 yuan, and the relative change rate of the leader's revenue in Round 5 was as follows. =0.23%≤1%, continue to meet the requirement.

[0112] Round 6: The leader's pricing remained unchanged, and the total follower response was 511 kWh; the leader's revenue was 12,810 yuan, and the relative change rate of the leader's revenue in Round 6 was as follows. =0.16%≤1%, and the condition is met for 3 consecutive rounds (rounds 4, 5, and 6), then convergence is declared.

[0113] In one embodiment, the scheduling device can intervene in oscillation scenarios through an arbitrator. If oscillation occurs, for example, the leader's pricing fluctuates between 0.9 and 1.1, the return fluctuation exceeds 15%, and it fails to converge within 5 rounds, then the arbitrator will automatically intervene. The arbitrator detected max( )>0.15 and iteration≥5.

[0114] The arbitrator calls `nash_bargaining_solver`, taking the leader's current offer and the willingness of each follower to respond (extracted from history) as input, and solves for the Nash bargaining equilibrium.

[0115] Forced output scheme: peak electricity price of 0.94 yuan / kWh, off-peak electricity price of 0.26 yuan / kWh, and capacity allocated to each follower proportionally.

[0116] The arbitrator broadcasts the message: "Game oscillation detected, forced convergence has been achieved. The final solution is: peak electricity price of 0.94 yuan / kWh, and off-peak electricity price of 0.26 yuan / kWh. Please execute this."

[0117] After the game converges, the security verification agent checks each control command one by one. The checks include: Power limiting: ≤ ,in, This represents the charging / discharging power command value to be sent to the energy storage device corresponding to the i-th follower agent, which is to be verified for safety. This indicates the maximum permissible power of the energy storage device as determined by the equipment parameter table. From the equipment parameter table.

[0118] SOC upper and lower limits: ≤ ≤ ,in =0.1 (unless the power-saving mode is increased to 0.6), It can be determined based on the operating parameters of the corresponding energy storage device.

[0119] Slope rate: ≤ , It is generally set at 20% / min of the rated power.

[0120] Non-negative returns: ≥0 and ≥0, to prevent vicious competition.

[0121] in, This represents the charge limit of the energy storage device corresponding to the i-th follower agent; This indicates the state of charge of the energy storage device at the beginning of the t-th scheduling period; This indicates the upper limit of the state of charge (SOC) of the energy storage device. The SOC is expressed as a dimensionless ratio; a higher value indicates a higher percentage of remaining energy in the storage device.

[0122] and Let represent the charging and discharging power of the energy storage device corresponding to the i-th follower agent during the t-th and t-1th scheduling periods, respectively; This indicates the magnitude of change in charging and discharging power between adjacent scheduling periods; The maximum allowable power change is determined by the product of the allowable power change rate and the time interval between adjacent scheduling periods. When the allowable power change rate is set to 20% / min of the rated power, the maximum allowable power change amplitude is calculated based on the actual time interval between adjacent scheduling periods.

[0123] This represents the cumulative gain of the i-th follower agent within the target scheduling period; This represents the cumulative reward of the leader agent within the same objective scheduling cycle; ≥0 and ≥0 indicates that the scheduling schemes corresponding to each follower agent and the leader agent should all satisfy the condition that the cumulative revenue is non-negative.

[0124] If an instruction violates a constraint, for example, a capacity follower plans to discharge 250kW, but its maximum power is only 200kW, then the safety verification agent calls the correction function: in, This indicates the candidate charge / discharge power command to be solved. This represents the original charging and discharging power command of the energy storage device corresponding to the i-th follower agent. The L2 norm represents the deviation between the candidate instruction and the original instruction, used to measure the magnitude of the correction. When the instruction is a power value for a single time period, the L2 norm equals the absolute value of the difference between the two; when the instruction is a power sequence consisting of multiple time periods, the L2 norm equals the square root of the sum of the squares of the power differences between the corresponding time periods. `argmin` represents the decision variable used to minimize the objective function. This indicates the revised power command. This represents the set of feasible power commands determined by constraints such as applicable power limits, state of charge limits, and ramp limits for the energy storage device corresponding to the i-th follower agent. The above correction process selects the command with the smallest deviation from the original command from the candidate commands that satisfy the applicable constraints as the correction result.

[0125] The corrected power is 200kW, with the following explanation: "Your commanded discharge power of 250kW exceeds the maximum allowable value of 200kW and has been automatically corrected to 200kW. Please confirm."

[0126] Based on the above embodiments, please refer to Figure 3 , Figure 3 This is a flowchart illustrating the third embodiment of the hybrid energy storage scheduling method based on LLM agents and game theory in this application. In this third embodiment, step S40 further includes: Step S401: Perform multi-layer wavelet packet decomposition on the total power command to obtain multiple frequency band power components, and determine the center frequency corresponding to each frequency band power component; Step S402: Compare the center frequency corresponding to each frequency band power component with the preset high-frequency threshold and the preset low-frequency threshold respectively to obtain the comparison results; Step S403: Generate a frequency band component allocation strategy based on the comparison results, and allocate each frequency band power component to the corresponding follower agent based on the allocation strategy. Each follower agent is configured to superimpose the allocated frequency band power component with the basic charge and discharge power curve of the corresponding energy storage type to generate a target scheduling instruction sequence.

[0127] It should be noted that the frequency band component allocation strategy includes: The power components of the frequency band with a center frequency higher than the preset high-frequency threshold are allocated to the power follower; The power components of the frequency band with a center frequency higher than the preset low-frequency threshold and lower than or equal to the preset high-frequency threshold are allocated to the capacity follower. The power components of the frequency band with a center frequency lower than or equal to the preset low-frequency threshold are allocated to the ultra-long-duration follower.

[0128] In practical implementation, the scheduling equipment uses total power commands. For the input signal, the Daubechies-4 wavelet basis is used to perform 4-level wavelet packet decomposition, resulting in 16 frequency band components: in, The net power response required by the energy storage system as defined in step one, in kW. This represents the k-th frequency band component after decomposition, with units of kW. The sum of all components equals the original signal and they do not overlap. Each component... With center frequency The unit is Hz. The system presets two frequency thresholds: =0.0167Hz (corresponding to a 1-minute cycle). =0.0000694Hz (corresponding to a period of 4 hours). According to The comparison with the threshold is assigned as follows: like > (High frequency, period < 1 minute), will Assigned to power-type followers (flywheels / supercapacitors) for real-time response.

[0129] like < ≤ (Low to medium frequency, cycle 1 minute - 4 hours), will Allocate capacity-type followers (lithium-ion batteries / flow batteries) to their charge / discharge plans.

[0130] like ≤ (Extremely low frequency, period > 4 hours), will The ultra-long-duration follower (compressed air / hydrogen) is assigned for long-duration scheduling.

[0131] Heat load component It is assigned separately to thermal energy storage followers (molten salt thermal energy storage) and does not participate in wavelet packet decomposition.

[0132] After the allocation is completed, each follower agent will receive The component is added to its original charging and discharging plan to form the final execution instruction. Threshold The strategy commander agent can dynamically adjust the frequency based on real-time grid frequency fluctuations: increase the frequency when fluctuations are severe. This allows more components to be included in the high-frequency range, enhancing the participation of power-type energy storage.

[0133] In some embodiments, each agent maintains an independent long-term memory (LLM) implemented using a vector database. Each memory entry contains: a state embedding vector (512-dimensional, encoded by Sentence-BERT to show the current system state), a decision action (pricing or power sequence), the reward after execution, and a timestamp. When the agent needs to make a decision, it first encodes the current state as a vector, retrieves the five most similar historical memories using cosine similarity, and injects the actions and rewards from these memories as contextual cues into the LLM to help it generate better decisions. This effectively reduces the LLM illusion and enables experience reuse.

[0134] In some embodiments, prior to deployment of the hybrid energy storage scheduling system, the leader agent and its followers are trained in a simulation environment through self-game adversarial training. The simulation environment simulates the randomness of electricity prices, load, renewable energy output, and the physical constraints of the energy storage devices. Training employs the Multi-Agent Deep Deterministic Policy Gradient (MADDPG) algorithm, collecting a large number of samples (states, joint actions, and joint rewards) after 5000 rounds of gameplay. Then, knowledge distillation is used to train a lightweight feedforward neural network (4 fully connected layers, 64 neurons per layer) to simulate the leader's pricing strategy. This lightweight model reduces inference time from approximately 400ms for the original LLM to less than 10ms, allowing for real-time deployment on edge devices; while the original LLM agent remains responsible for long-term planning (minutes or more), human-computer interaction, and decision-making in complex scenarios.

[0135] In some embodiments, at the end of each scheduling cycle (e.g., 24 hours), each follower agent writes its actual execution data (charge / discharge amount, SOC change, actual revenue, and decay amount) into its own long-term memory. The strategy commander agent triggers a daily reflection at dawn: it calls the LLM to analyze the previous day's scheduling performance, generates improvement suggestions (e.g., "during the midday peak solar power generation period, flywheel participation should be increased"), and stores these suggestions in the knowledge base for subsequent scheduling via RAG retrieval. Simultaneously, the macro-level experience of the entire cycle (initial state, action sequence, total revenue, and final state) is stored in a separate "evolutionary experience buffer." During offline self-game training, these macro-level experiences are sampled to fine-tune the agent's policy network, enabling the system to continuously adapt to long-term drift such as equipment aging and changes in electricity pricing policies.

[0136] Reference Figure 4 , Figure 4 This is a schematic diagram of the interaction timing of various intelligent agents in one embodiment of this application. Figure 4 The interaction process between the strategy commander agent, the leader agent, the capacity-type follower agent, the power-type follower agent, the ultra-long-duration follower agent, the heterogeneous resource arbitrator agent, and the security verification agent is illustrated. Figure 4The vertical dashed lines in the diagram indicate that the corresponding agents are continuously in an interactive state during the scheduling cycle, the horizontal arrows indicate the direction of message passing between agents, and k represents the current master-slave game iteration round.

[0137] The strategy commander agent parses the natural language scheduling instructions input by the user, obtaining a set of scheduling parameters including quantization constraints and target weights. It then combines this with preprocessed multi-source heterogeneous data to generate or obtain a total power instruction. The strategy commander agent sends the total power instruction and scheduling parameter set to the leader agent, and sends the quantization constraints and target weights related to each energy storage type to the capacity-type follower agent, power-type follower agent, and ultra-long-duration follower agent, respectively, enabling each agent to execute subsequent decisions based on a consistent scheduling objective.

[0138] The leader agent constructs an optimization model based on the total power command, scheduling parameter set, historical electricity price sequence, and the follower response results from the previous round. It then generates the k-th round policy message by invoking an electricity price prediction tool and an optimization solver. The policy message can include electricity price strategy, energy storage capacity leasing strategy, power demand, response time period, and task constraints matching the energy storage type. The leader agent sends the k-th round policy message to each follower agent participating in the master-slave game.

[0139] After receiving the strategy message in the k-th round, each follower agent executes a ReAct decision loop based on the operating status of the corresponding energy storage device. Capacity-type follower agents determine the response capacity and charge / discharge power curves based on state of charge, health status, available capacity, charge / discharge power limits, and lifetime degradation costs. Power-type follower agents determine the high-frequency power response and operating status based on the available power, response speed, and operating temperature of the flywheel or supercapacitor. Long-duration follower agents determine the long-cycle charge / discharge power curves based on energy storage pressure, remaining energy, conversion efficiency, weather forecast data, and load forecast data. Each follower agent then returns its decision results to the leader agent.

[0140] The leader agent calculates the current round's payoff based on the response capacity and charging / discharging power curves returned by each follower agent, and sends the sum of the current round's payoff and follower responses to the strategy commander agent. The strategy commander agent determines whether the master-slave game has converged based on the changes in leader payoff and follower strategies between adjacent iterations. If the preset convergence condition is not met, the strategy commander agent sends a strategy update message to the leader agent, which then adjusts its electricity pricing strategy and energy storage capacity leasing strategy before proceeding to the next iteration.

[0141] When the strategy commander agent detects continuous oscillations in the leader's payoff, pricing results exceeding a preset reasonable range, or insufficient follower response capacity, it sends the corresponding abnormal status to the heterogeneous resource arbitrator agent. The heterogeneous resource arbitrator agent, based on the payoffs, threat points, available capacity, and energy storage operation constraints of each participating entity, invokes a Nash bargaining solver or an alternating direction multiplier method solver to generate a arbitrated pricing scheme and task allocation scheme. The arbitrated pricing is then sent to the leader agent, while capacity allocation tasks, high-frequency response tasks, and long-cycle response tasks are sent to the corresponding follower agents.

[0142] Under normal convergence of the master-slave game, the strategy commander agent sends the converged scheduling scheme to the safety verification agent. When the heterogeneous resource arbitrator agent intervenes, it sends the arbitrated scheduling scheme to the safety verification agent. The safety verification agent checks whether each scheduling instruction meets safety constraints such as charging / discharging power limits, energy storage state limits, equipment temperature limits, ultra-long-term energy storage pressure limits, and prohibition of simultaneous charging / discharging. For scheduling instructions that meet the safety constraints, the safety verification agent generates a verification pass result; for scheduling instructions that do not meet the safety constraints, the safety verification agent corrects them through power limiting, task reallocation, or backup energy storage switching, and feeds back the verification result and correction instructions to the strategy commander agent.

[0143] After security verification, the target scheduling command is sent to the controller of the corresponding energy storage device to control each energy storage device to perform charging or discharging operations. After the scheduling cycle ends, the actual charging and discharging amount, energy storage state changes, actual benefits, and equipment degradation of each energy storage device are written into the long-term memory of the corresponding intelligent agent as execution feedback, providing historical data for retrieval enhancement decisions and strategy updates in subsequent scheduling cycles.

[0144] It should be noted that the embodiments of this application include the following beneficial effects: 1. A Stackelberg master-slave game structure was constructed, consisting of a leader LLM agent (energy storage operator) and at least two follower LLM agents belonging to different technology types (power type, capacity type, ultra-long-duration type, thermal energy storage). The leader agent autonomously formulates and publishes price / lease strategies in natural language form through LLM inference (first move); each follower agent uses the ReAct paradigm (think-act-observe) to call its own tools (SOC prediction, lifetime decay model, optimization solver) to autonomously decide the optimal charging and discharging power and respond in natural language (last move); the strategy commander agent monitors the game iteration process in real time and makes real-time convergence judgments; each agent is equipped with an independent long-term memory, tool call interface, and natural language communication capability. Compared with existing technologies, existing multi-agent schemes mostly involve equal negotiation or are based on fixed rules, lacking a hierarchical interest game relationship between leader and follower, and the large language model is only used as a text generation or passive coordination tool, lacking autonomous perception, planning, memory, tool call, and natural language interaction capabilities. This application is the first to deeply integrate LLM agents with Stackelberg game theory, forming a clear first-move-last-move hierarchical decision-making mechanism.

[0145] Beneficial effects: It achieves optimal economic allocation among multiple stakeholders, solving the problem of irreconcilable conflicts of interest in centralized scheduling; natural language communication makes the system readable and explainable by both humans and machines, allowing maintenance personnel to directly view dialogue logs and understand decision-making logic; the system is highly scalable, requiring only modification of role prompts when adding new energy storage types, without reprogramming; and the game convergence condition avoids subjective judgment.

[0146] 2. In this application, all communication between LLM agents adopts a dynamically generated natural language dialogue format. The LLM generates message content in real time based on the context, without the need to predefine message types or field structures. Compared with existing technologies, traditional multi-agent systems rely on predefined structured message formats (such as JSON, binary protocols) for communication. This fixed mode is difficult to cope with unforeseen interaction scenarios, and the communication protocol must be modified when adding new roles. This application completely entrusts the dynamic generation of communication content by the LLM. The communication protocol only specifies the underlying message passing method and does not constrain the structure of the message content.

[0147] Beneficial effects: The system's scalability is greatly improved. Any new energy storage type only needs to have its role prompt words defined, and the communication protocol does not need to be modified. Anyone can intervene at any time to view the agent's dialogue logs and clearly understand the entire process of bidding, response, negotiation, and arbitration. The dialogue logs have inherent fault diagnosis capabilities and can be directly used to review the reasons for game failures. Agents can handle undefined abnormal situations (such as a follower reporting "My SOC is too low to discharge"), while traditional structured protocols require enumerating all possible states in advance.

[0148] 3. In this application, each LLM agent adopts the ReAct paradigm (think → act → observe loop) for decision-making. The agent first generates internal reasoning text based on its current state and memory, then autonomously decides to call external tools (electricity price prediction, SOC prediction, optimization solver, equipment control interface, etc.), and continues reasoning based on the numerical results returned by the tools until the final action command is generated. Compared with existing technologies, the scheduling logic of traditional energy storage EMS is based on hard-coded rules or fixed models trained offline. Once deployed, it cannot adapt to changes in electricity price policies, equipment aging, or the addition of new equipment, requiring reprogramming or retraining. This application does not rely on predefined decision trees, but allows the LLM agent to autonomously plan action sequences online.

[0149] Beneficial effects: The system has zero-sample adaptation capability. When faced with new electricity pricing policies, the access of new energy storage devices, or unforeseen operating scenarios, it can generate reasonable strategies through LLM general reasoning without reprogramming or training; it greatly reduces the long-term operation and maintenance costs of the system and avoids the problem of rule base explosion; the decision-making process of the agent is interpretable - the internal thinking text can be recorded and reviewed, which facilitates debugging and optimization.

[0150] 4. This application maintains an independent long-term memory (vector database) for each agent. Each memory entry contains a state embedding vector, a decision action, an execution reward, and a timestamp. When the agent needs to make a decision, it first encodes the current state as a vector, retrieves the most similar historical memories using cosine similarity, and injects the actions and rewards from these memories as contextual cues into the LLM to guide the model to refer to past successful experiences. Compared with existing technologies, current LLM agent systems typically use a single shared memory, leading to mutual interference between experiences of different types of energy storage (power type and capacity type); and decision-making relies only on the current context, unable to actively invoke historical cases to assist in decision-making. This application uses an independent memory to isolate experiences of different energy storage types and actively retrieves the most similar memories to inject cues.

[0151] Beneficial effects: Significantly reduces the "illusion" caused by LLM's lack of experience in specific fields, improving the accuracy and consistency of decision-making; experience of different types of energy storage is isolated from each other, avoiding negative transfer effects (such as incorrectly applying experience of lithium batteries to flywheels); system performance continuously improves itself during long-term operation, and historical successful experience can be repeatedly reused; the retrieval process can be quantified (cosine similarity), and the effectiveness of memory entries can be screened and eliminated through benefit feedback.

[0152] 5. This application employs wavelet packet decomposition to decompose the total power command into multiple frequency band components. Based on a comparison of the center frequency of each component with a preset threshold (a 1-minute period for high-frequency components and a 4-hour period for extremely low-frequency components), high-frequency components are automatically allocated to power-type followers, medium- and low-frequency components to capacity-type followers, and extremely low-frequency components to ultra-long-duration followers. The thresholds can be dynamically adjusted by the strategy commander based on real-time grid frequency fluctuations. Compared to existing technologies, most existing hybrid energy storage scheduling methods employ low-pass filtering or empirical allocation rules with fixed time constants, failing to dynamically adjust the allocation ratios of power-type, capacity-type, and ultra-long-duration energy storage based on real-time grid fluctuation characteristics. This can lead to high-frequency fluctuations being incorrectly allocated to slow energy storage (resulting in delayed response) or low-frequency components being allocated to fast energy storage (resulting in wasted efficiency). This application is based on rigorous signal decomposition theory, and the thresholds can be dynamically adjusted.

[0153] Beneficial effects: It fully leverages the time-scale advantages of different energy storage technologies—power-type energy storage only handles millisecond to second-level fluctuations, capacity-type energy storage handles minute to hour-level energy transfers, and ultra-long-duration energy storage handles trend components for more than 4 hours; it avoids the problems of high-frequency fluctuations being misallocated to slow energy storage, resulting in untimely response, or low-frequency components being allocated to fast energy storage, resulting in inefficiency; the overall energy storage system utilization efficiency is greatly improved; the threshold dynamic adjustment mechanism enables the system to adapt to the severity of grid frequency fluctuations, automatically assigning more tasks to power-type energy storage when the grid is unstable.

[0154] 6. Before system deployment, this application conducts self-game adversarial training between the leader agent and each follower agent in a simulation environment (using a multi-agent deep deterministic policy gradient algorithm) to collect a large number of samples (states, joint actions, joint rewards). Then, using knowledge distillation, a lightweight feedforward neural network (4 fully connected layers) is trained as the student model, with the policy output of the original LLM as the teacher signal. The lightweight model is deployed on an edge device for real-time response; the original LLM agent is still responsible for long-term planning (minutes or more), human-computer interaction, and complex scenario decision-making. Compared with existing technologies, LLM inference latency is high and cannot be directly used for millisecond-level real-time control scenarios, while simple model quantization or pruning will severely reduce decision accuracy. This application generates training data through self-game and then obtains a lightweight model through knowledge distillation, realizing the collaborative work of the LLM agent and the fast-response model.

[0155] Beneficial effects: While maintaining high decision-making accuracy, inference latency was reduced from hundreds of milliseconds to milliseconds, achieving for the first time the engineering deployment of LLM agents in high real-time energy storage control; the system achieves a division of labor between intelligence and real-time performance—fast response is handled by the lightweight model, while complex planning is handled by the LLM, without interference between them; self-game training enables the lightweight model to cover the decision-making patterns of the LLM in a large number of scenarios, and the distillation process ensures that the output distribution of the student model and the teacher model is as close as possible.

[0156] 7. This application incorporates a heterogeneous resource arbitrator agent that continuously monitors the game process. When oscillations are detected (leader payoff fluctuations exceed a threshold and fail to converge within multiple rounds), leader pricing exceeds the historical reasonable range, or the total follower response capacity is lower than the leader's expected capacity, the arbitrator automatically intervenes, invoking the Nash bargaining solver or the alternating direction multiplier method. Considering the threat points of each party (payoffs when not cooperating), it forcibly generates a set of Pareto-improved pricing and capacity allocation schemes, which are then broadcast to all agents in natural language for execution. Compared to existing technologies, existing schemes typically require manual intervention to adjust parameters or restart the system when game oscillations or non-convergence occur, lacking an automated conflict resolution mechanism. This application designs the arbitrator as an independent LLM agent capable of understanding the semantics of game conflict and invoking mathematical optimization methods to force convergence.

[0157] Beneficial effects: The system is robust and self-healing, capable of automatically detecting and resolving game conflicts, avoiding infinite oscillations or deadlocks; it significantly reduces the number of manual maintenance interventions, lowering operating costs; the generated Pareto improvement schemes ensure that the benefits of each party are at least no less than the level under non-cooperation, avoiding unilateral loss of interests due to forced mediation; the arbitrator's intervention conditions and algorithms are transparent and configurable, facilitating adjustments according to different application scenarios.

[0158] This application also provides a hybrid energy storage scheduling system based on LLM agents and game theory. Please refer to [link / reference]. Figure 5 , Figure 5 This is a schematic diagram of the module structure of a hybrid energy storage scheduling system based on LLM agents and game theory. The system includes: The data processing module 10 is used to preprocess the collected multi-source heterogeneous data and generate a total power command based on the preprocessed multi-source heterogeneous data. The instruction response module 20 is used to respond to a natural language scheduling instruction sent by the user and generate a scheduling parameter set based on the natural language scheduling instruction. The scheduling parameter set includes quantization constraints and target weights. The master-slave game module 30 is used to perform master-slave game iterations through a leader agent and multiple follower agents based on the total power command and the scheduling parameter set, to obtain the basic charge and discharge power curve in the converged state. The leader agent and the multiple follower agents all use a large language model as the inference engine. The multiple follower agents correspond to different types of energy storage devices. The follower agents include capacity-type followers, power-type followers and ultra-long-duration followers. The frequency band decomposition module 40 is used to decompose the total power command into frequency bands to obtain multiple frequency band power components, and generate a target scheduling command sequence based on the multiple frequency band power components and the basic charge and discharge power curve. The energy storage scheduling module 50 is used to send the target scheduling instruction sequence to the controller of each energy storage device, so that each controller controls the corresponding energy storage device to perform charging and discharging operations according to the target scheduling instruction sequence.

[0159] The hybrid energy storage scheduling system based on LLM agents and game theory provided in this application, employing the hybrid energy storage scheduling method based on LLM agents and game theory in the above embodiments, can solve the technical problems to be addressed by the aforementioned hybrid energy storage scheduling method based on LLM agents and game theory. Compared with the prior art, the beneficial effects of the hybrid energy storage scheduling system based on LLM agents and game theory provided in this application are the same as those of the hybrid energy storage scheduling method based on LLM agents and game theory provided in the above embodiments, and other technical features of the hybrid energy storage scheduling system based on LLM agents and game theory are the same as those disclosed in the methods of the above embodiments, and will not be repeated here.

[0160] This application provides a hybrid energy storage scheduling device based on LLM agents and game theory. The hybrid energy storage scheduling device based on LLM agents and game theory includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to execute the hybrid energy storage scheduling method based on LLM agents and game theory in any of the above embodiments.

[0161] The following is for reference. Figure 6This document illustrates a structural schematic diagram of a hybrid energy storage scheduling device based on LLM agents and game theory, suitable for implementing embodiments of this application. The hybrid energy storage scheduling device based on LLM agents and game theory in this application embodiment may include, but is not limited to, mobile terminals such as mobile phones, laptops, digital broadcast receivers, PDAs (Personal Digital Assistants), PADs (Portable Android Devices), PMPs (Portable Media Players), and in-vehicle terminals (e.g., in-vehicle navigation terminals), as well as fixed terminals such as digital TVs and desktop computers. Figure 6 The hybrid energy storage scheduling device based on LLM agents and game theory shown is merely an example and should not impose any limitations on the functionality and scope of use of the embodiments of this application.

[0162] like Figure 6 As shown, the hybrid energy storage scheduling device based on LLM agents and game theory may include a processing unit 1001 (e.g., a central processing unit, a graphics processing unit, etc.), which can perform various appropriate actions and processes according to programs stored in ROM 1002 (Read Only Memory) or programs loaded from storage device 1003 into RAM 1004 (Random Access Memory). RAM 1004 also stores various programs and data required for the operation of the hybrid energy storage scheduling device based on LLM agents and game theory. The processing unit 1001, ROM 1002, and RAM 1004 are interconnected via bus 1005. I / O interface 1006 is also connected to the bus. Typically, the following systems can be connected to I / O interface 1006: input devices 1007 including, for example, touchscreens, touchpads, keyboards, mice, image sensors, microphones, accelerometers, gyroscopes, etc.; output devices 1008 including, for example, LCDs (Liquid Crystal Displays), speakers, vibrators, etc.; storage devices 1003 including, for example, magnetic tapes, hard disks, etc.; and communication devices 1009. Communication device 1009 allows the hybrid energy storage scheduling device based on LLM agents and game theory to communicate wirelessly or wiredly with other devices to exchange data. Although the figure shows a hybrid energy storage scheduling device based on LLM agents and game theory with various systems, it should be understood that it is not required to implement or possess all the systems shown. More or fewer systems can be implemented alternatively.

[0163] The hybrid energy storage scheduling device based on LLM agents and game theory provided in this application, employing the hybrid energy storage scheduling method based on LLM agents and game theory in the above embodiments, can solve the technical problems to be solved by the aforementioned hybrid energy storage scheduling method based on LLM agents and game theory. Compared with the prior art, the beneficial effects of the hybrid energy storage scheduling device based on LLM agents and game theory provided in this application are the same as the beneficial effects of the hybrid energy storage scheduling method based on LLM agents and game theory provided in the above embodiments, and other technical features in this hybrid energy storage scheduling device based on LLM agents and game theory are the same as those disclosed in the previous embodiment method, and will not be repeated here.

[0164] It should be understood that the various parts disclosed in this application can be implemented using hardware, software, firmware, or a combination thereof. In the description of the above embodiments, specific features, structures, materials, or characteristics can be combined in any suitable manner in one or more embodiments or examples.

[0165] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0166] This application provides a computer-readable storage medium having computer-readable program instructions (i.e., a computer program) stored thereon, which are used to execute the hybrid energy storage scheduling method based on LLM agents and game theory in the above embodiments.

[0167] The computer-readable storage medium provided in this application can be a tangible medium capable of storing computer programs, such as a hard disk, random access memory, read-only memory, flash memory, optical storage device, or magnetic storage device. When the computer program is executed by a processor, it implements the above-described hybrid energy storage scheduling method based on LLM agents and game theory.

[0168] The aforementioned computer-readable storage medium may be included in a hybrid energy storage scheduling device based on LLM agents and game theory; or it may exist independently and not be assembled into a hybrid energy storage scheduling device based on LLM agents and game theory.

[0169] The aforementioned computer-readable storage medium carries one or more programs. When the aforementioned one or more programs are executed by the hybrid energy storage scheduling device based on LLM agents and game theory, the hybrid energy storage scheduling device based on LLM agents and game theory performs the hybrid energy storage scheduling method based on LLM agents and game theory as described in the above embodiments.

[0170] Computer program code for performing the operations of this application can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, as well as conventional procedural programming languages ​​such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including LAN (Local Area Network) or WAN (Wide Area Network)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0171] The modules described in the embodiments of this application can be implemented in software or hardware. The names of the modules do not necessarily limit the functionality of the unit itself.

[0172] The readable storage medium provided in this application is a computer-readable storage medium that stores computer-readable program instructions (i.e., a computer program) for executing the aforementioned hybrid energy storage scheduling method based on LLM agents and game theory, and is capable of solving the technical problems to be solved by the aforementioned hybrid energy storage scheduling method based on LLM agents and game theory. Compared with the prior art, the beneficial effects of the computer-readable storage medium provided in this application are the same as the beneficial effects of the hybrid energy storage scheduling method based on LLM agents and game theory provided in the above embodiments, and will not be repeated here.

[0173] The above description is only a part of the embodiments of this application and does not limit the patent scope of this application. All equivalent structural transformations made under the technical concept of this application and using the contents of the specification and drawings of this application, or direct / indirect applications in other related technical fields, are included in the patent protection scope of this application.

Claims

1. A hybrid energy storage scheduling method based on LLM agents and game theory, characterized in that, The method includes: The collected multi-source heterogeneous data is preprocessed, and a total power command is generated based on the preprocessed multi-source heterogeneous data. In response to a natural language scheduling instruction sent by a user, a scheduling parameter set is generated based on the natural language scheduling instruction, the scheduling parameter set including quantization constraints and target weights; Based on the total power command and the scheduling parameter set, the basic charge and discharge power curve in the converged state is obtained by performing master-slave game iteration through the leader agent and multiple follower agents. The leader agent and multiple follower agents all use a large language model as the inference engine. The multiple follower agents correspond to different types of energy storage devices. The follower agents include capacity followers, power followers and ultra-long time followers. The total power command is decomposed into multiple frequency band power components, and a target scheduling command sequence is generated based on the multiple frequency band power components and the basic charge and discharge power curve. The target scheduling instruction sequence is sent to the controller of each energy storage device, so that each controller controls the corresponding energy storage device to perform charging and discharging operations according to the target scheduling instruction sequence.

2. The method as described in claim 1, characterized in that, The process of obtaining the basic charge-discharge power curve in the convergent state by performing master-slave game iterations through a leader agent and multiple follower agents based on the total power command and the scheduling parameter set includes: The leader agent is invoked to generate the electricity price and leasing strategy message for the current round based on the total power command and the scheduling parameter set; The electricity price and leasing strategy message for the current round is broadcast to each follower agent participating in the master-slave game; Each follower agent is invoked to make a charging and discharging decision for the current round based on the electricity price and leasing strategy message of the current round, the operating status of the corresponding energy storage device, and the scheduling parameter set. A follower response message is generated based on the charging and discharging decision result. The follower response message includes the response capacity and the response charging and discharging power curve. The follower response message is sent to the leader agent, which then calculates the leader's reward for the current round based on the follower response message. The electricity price and leasing strategy messages are then updated based on the leader's reward for the current round. If the preset convergence condition is not met, proceed to the next round of master-slave game iteration based on the updated electricity price and leasing strategy messages; if the preset convergence condition is met, obtain the basic charging and discharging power curve based on the follower response messages of the current round.

3. The method as described in claim 2, characterized in that, The leader agent is configured to perform the following steps during the master-slave game iteration process: Historical electricity price sequences are extracted from preprocessed multi-source heterogeneous data, and the historical electricity price sequences are input into a pre-trained electricity price prediction model to obtain the predicted electricity price sequence within the target scheduling period. A leader optimization model is constructed based on the predicted electricity price sequence, the total power command, and the scheduling parameter set. The optimization objective of the leader optimization model is to maximize the leader's revenue. The constraints of the leader optimization model include the state of charge constraint and the charging and discharging power constraint of the leader's energy storage device. Using the historical response messages of the follower agent as initial optimization parameters, the mixed-integer linear programming solver is invoked to solve the leader optimization model, obtaining the solution status identifier, candidate electricity price strategy, candidate energy storage capacity leasing strategy and corresponding leader revenue; If the solution status indicator indicates a successful solution, a target electricity price strategy and a target energy storage capacity leasing strategy are selected from the candidate electricity price strategies and the candidate energy storage capacity leasing strategies based on the leader's reward. An electricity price and leasing strategy message is generated and broadcast to each follower agent participating in the master-slave game.

4. The method as described in claim 3, characterized in that, The objective function of the leader optimization model is: in, The subscript represents the leader's benefit to the leader agent. Represents the leader intelligent agent; Indicates the total number of scheduling periods; Indicates the duration of a single scheduling period; For the leader agent in the 1st The electricity price set for each scheduling period; Indicates the sequence number of the scheduling period; Represents the set of follower agents; i represents the index of the follower agent. Let be the discharge power of the i-th follower during the t-th scheduling period; This represents the unit discharge operating cost of the leader's own energy storage device; For rental rates; Capacity leased for follower agents; This represents the discharge amount of the leader agent's own energy storage device during the t-th scheduling period.

5. The method as described in claim 2, characterized in that, The capacity-type follower is configured to perform the following steps during the master-slave game: The electricity price and leasing strategy messages of the current round are parsed to obtain the target electricity price strategy and target energy storage capacity leasing strategy of the current round, and the energy storage type identifier and operating status of the corresponding energy storage device are extracted from the preprocessed multi-source heterogeneous data. A follower optimization model is constructed based on the target electricity price strategy, the target energy storage capacity leasing strategy, the energy storage type identifier, and the operating status. The optimization objective of the follower optimization model is to maximize the follower's revenue. The constraints of the follower optimization model include charging and discharging power constraints, available capacity constraints of energy storage devices, energy storage state constraints, and energy storage state transition constraints. Using the historical response messages of the corresponding follower agent as the initial optimization parameters, the local optimizer that matches the energy storage type of the corresponding energy storage device is called to solve the follower optimization model, and the solution status identifier, candidate charging and discharging strategies, candidate response capacity, and follower benefits corresponding to each candidate charging and discharging strategy are obtained. If the solution status indicator indicates that the solution is successful, a target charging and discharging strategy is selected from the candidate charging and discharging strategies based on the follower's reward, and a follower response message is generated according to the target charging and discharging strategy and the corresponding candidate response capacity. The follower response message is then sent to the leader agent.

6. The method as described in claim 5, characterized in that, The objective function of the follower optimization model is: in, Indicates the first The cumulative gain of a follower agent within a scheduling cycle; Indicates the first The energy storage device corresponding to each follower agent is in the first... Charging power during each scheduling period; Indicates the first The unit lifetime degradation cost coefficient of the energy storage device corresponding to each follower agent. Indicates the first The energy storage device corresponding to each follower agent is in the first... The charging and discharging power during each scheduling period.

7. The method as described in claim 2, characterized in that, The ultra-long-duration follower is configured to perform the following steps during the master-slave game: Extract the target electricity price strategy and target energy storage capacity leasing strategy for the current round from the current round's electricity price and leasing strategy message, and obtain the energy storage pressure range, charge and discharge power range, and energy conversion efficiency of the ultra-long-duration energy storage device; Meteorological forecast data and load forecast data for a preset planning period are extracted from the preprocessed multi-source heterogeneous data; The energy storage pressure range and charge / discharge power range are discretized to obtain multiple pressure states and multiple charge / discharge power levels. Based on the multiple pressure states, the multiple charging and discharging power levels, the energy conversion efficiency, and the scheduling parameter set, a state transition relationship and a set of ultra-long-term scheduling constraints are constructed. Based on the state transition relationship and the ultra-long-term scheduling constraint set, the charging and discharging revenues corresponding to the target electricity price strategy and the target energy storage capacity leasing strategy are cumulatively solved to obtain the target long-term charging and discharging power sequence. Based on the target long-cycle charge-discharge power sequence, the response capacity and response charge-discharge power curve of the ultra-long-duration energy storage device are determined, and a follower response message is generated and sent to the leader agent.

8. The method as described in claim 1, characterized in that, The step of performing frequency band decomposition on the total power command to obtain multiple frequency band power components, and generating a target scheduling command sequence based on the multiple frequency band power components and the basic charge-discharge power curve, includes: The total power command is decomposed into multiple wavelet packets to obtain multiple frequency band power components, and the center frequency corresponding to each frequency band power component is determined. The center frequency corresponding to each power component of the frequency band is compared with the preset high-frequency threshold and the preset low-frequency threshold respectively to obtain the comparison results; Based on the comparison results, a frequency band component allocation strategy is generated, and each frequency band power component is allocated to the corresponding follower agent based on the allocation strategy. Each follower agent is configured to superimpose the allocated frequency band power component with the basic charge and discharge power curve of the corresponding energy storage type to generate a target scheduling instruction sequence. The frequency band component allocation strategy includes: The power components of the frequency band with a center frequency higher than the preset high-frequency threshold are allocated to the power follower; The power components of the frequency band with a center frequency higher than the preset low-frequency threshold and lower than or equal to the preset high-frequency threshold are allocated to the capacity follower. The power components of the frequency band with a center frequency lower than or equal to the preset low-frequency threshold are allocated to the ultra-long-duration follower.

9. A hybrid energy storage scheduling system based on LLM agents and game theory, applying the hybrid energy storage scheduling method based on LLM agents and game theory as described in any one of claims 1 to 8, characterized in that, The system includes: The data processing module is used to preprocess the collected multi-source heterogeneous data and generate a total power command based on the preprocessed multi-source heterogeneous data. The instruction response module is used to respond to the natural language scheduling instruction sent by the user and generate a scheduling parameter set based on the natural language scheduling instruction. The scheduling parameter set includes quantization constraints and target weights. The master-slave game module is used to obtain the basic charge and discharge power curve in the converged state by performing master-slave game iteration through a leader agent and multiple follower agents based on the total power command and the scheduling parameter set. The leader agent and the multiple follower agents all use a large language model as the inference engine. The multiple follower agents correspond to different types of energy storage devices. The follower agents include capacity followers, power followers and ultra-long time followers. The frequency band decomposition module is used to decompose the total power command into multiple frequency band power components, and generate a target scheduling command sequence based on the multiple frequency band power components and the basic charge and discharge power curve. The energy storage scheduling module is used to send the target scheduling instruction sequence to the controller of each energy storage device, so that each controller controls the corresponding energy storage device to perform charging and discharging operations according to the target scheduling instruction sequence.

10. A hybrid energy storage scheduling device based on LLM intelligent agents and game theory, characterized in that, The LLM agent-game-based hybrid energy storage scheduling device includes: a memory, a processor, and an LLM agent-game-based hybrid energy storage scheduling program stored in the memory. The processor is used to run the LLM agent-game-based hybrid energy storage scheduling program, which is configured to implement the LLM agent-game-based hybrid energy storage scheduling method as described in any one of claims 1 to 8.

Images