A flexible optimization scheduling method for integrated energy systems based on a large language model
By constructing a multi-source heterogeneous knowledge base and a rule-based logic encoding system, and training a large language model agent, flexible and optimized scheduling of integrated energy systems can be achieved. This solves the problem of insufficient adaptability of traditional methods, improves the flexibility and adaptability of scheduling, and supports a new energy system that is efficient and low-carbon.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2025-05-22
- Publication Date
- 2026-03-27
AI Technical Summary
Traditional integrated energy system optimization and scheduling methods are difficult to adapt to dynamically changing system states and market environments. They require real-time response and flexible adjustment, and existing methods are time-consuming and labor-intensive, making it difficult to integrate real-time market fluctuations and equipment status changes.
A flexible optimization scheduling method for integrated energy systems based on a large language model is adopted. By constructing a multi-source heterogeneous knowledge base and a rule logic encoding system, a large language model agent is trained to achieve dynamic reasoning and incremental learning, supporting natural language interaction and real-time scheduling strategy adjustment.
It enhances the flexibility and adaptability of the scheduling process, enabling rapid response to changes in user needs, ensuring the synchronous evolution of scheduling strategies with environmental changes, avoiding strategy failure, and supporting the construction of a new energy system that is efficient and low-carbon.
Smart Images
Figure CN120611899B_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application belongs to the field of comprehensive energy system optimal scheduling, and particularly relates to a comprehensive energy system flexible optimal scheduling method based on a large language model. BACKGROUND
[0002] Under the promotion of global energy transformation and the "double carbon" target, comprehensive energy systems, by integrating electricity, gas, heat and other forms of energy, have become an important carrier for improving energy utilization efficiency and promoting renewable energy consumption. Comprehensive energy systems are conducive to realizing multi-energy complementation and cascade utilization, such as generating electricity with industrial waste heat and supplying energy to buildings with photovoltaic and energy storage. With the rapid increase in the proportion of fluctuating power sources such as wind power and photovoltaic power, and the growing demand for energy cost and carbon emission control in factories and parks, the large-scale application of comprehensive energy systems has become a trend. At present, more than 60% of industrial parks in China have gradually introduced comprehensive energy services, and their low-carbon and flexible operation mode has provided key support for energy structure transformation.
[0003] Optimal scheduling of comprehensive energy systems is a key link in energy saving and carbon reduction. By coordinating the output of different energy equipment (such as gas turbines, energy storage systems, heat pumps, etc.), optimal scheduling can significantly reduce system operating costs, while reducing fossil energy consumption and carbon emissions. For example, storing electricity during low-price periods and discharging during peak periods, or preferentially scheduling low-carbon equipment when carbon prices rise, can achieve a win-win situation in terms of economy and environmental protection. However, with the increasing complexity of the system and the improvement of carbon market mechanisms and power spot market reforms, the fluctuations of energy prices, carbon prices and policy constraints are becoming more frequent, and traditional static optimization strategies are difficult to adapt to dynamically changing system states and market environments, so there is an urgent need for a scheduling method that can respond in real time and adjust flexibly.
[0004] Current comprehensive energy system optimal scheduling methods still have obvious limitations, including: (1) Most optimization models use fixed mathematical formulas and preset conditions, and once the user needs to adjust the time scale of scheduling (such as changing from day-ahead scheduling to intra-day real-time optimization) or add or remove equipment (such as temporarily connecting energy storage equipment), the model often needs to be rebuilt, which is time-consuming and laborious; (2) Existing methods usually rely on historical data for parameter setting, and are difficult to integrate real-time market fluctuations, policy adjustments or equipment state changes, leading to a gradual disconnection between the scheduling strategy and actual needs during use. SUMMARY
[0005] To solve the problems existing in the current energy system optimal scheduling method, the present application provides a comprehensive energy system flexible optimal scheduling method based on a large language model, which can create a flexible optimal scheduling scenario according to user needs and continuously update the agent's cognition, thereby improving the flexibility and adaptability of the optimal scheduling process, and contributing to the construction of an efficient and low-carbon new energy system in China.
[0006] The application adopts the following technical solutions:
[0007] A comprehensive energy system flexible optimization scheduling method based on a large language model, comprising the following steps:
[0008] Step S1, constructing a comprehensive energy system multi-source heterogeneous knowledge base and a rule logic coding system. Step S1 specifically comprises the following steps:
[0009] Step S11, constructing a comprehensive energy system multi-source heterogeneous knowledge base; specifically, integrating comprehensive energy system equipment technical parameters, system topological structure, transaction mechanism rules, policy texts, equipment and load historical data, historical energy prices and carbon prices, and simultaneously storing and associating mapping structured data (specifically including comprehensive energy system equipment technical parameters, system topological structure, equipment and load historical data, historical energy prices and carbon prices) and unstructured data (specifically including transaction mechanism rules and policy texts), establishing a comprehensive energy system multi-source heterogeneous knowledge base, which supports semantic retrieval. Wherein, semantic retrieval specifically refers to realizing intelligent association matching of user statements and stored data in the knowledge base by analyzing the context semantics of user natural language input sentences, returning structured parameters corresponding to the semantic meaning, rather than relying on traditional keyword matching. Figure One
[0010] Step S12, constructing a rule logic coding system; specifically, performing semantic analysis and entity relationship extraction on unstructured data in the comprehensive energy system multi-source heterogeneous knowledge base, identifying operation rules, constraint conditions and policy provisions in the comprehensive energy system field knowledge, and converting them into computer executable formal logic expressions (standardized expressions formed by mathematical symbols and logical operators, which can be directly called and executed by computer programs); through a dynamic mapping mechanism, binding operation rules with system real-time state variables, generating a set of computable constraint conditions (constraint conditions in the constraint condition set are in the form of computer executable formal logic expressions) suitable for different scheduling scenarios, and fusing the set of computable constraint conditions with structured data in the comprehensive energy system multi-source heterogeneous knowledge base, obtaining the rule logic coding system; the rule logic coding system supports multi-dimensional logical reasoning. Wherein, multi-dimensional logical reasoning refers to establishing an association mechanism and priority mechanism between different dimensions of rules, so as to analyze from multiple angles and weigh different possibilities under complex scheduling instructions, and finally generate a coherent and reasonable conclusion.
[0011] Step S2, based on the comprehensive energy system multi-source heterogeneous knowledge base and the rule logic coding system, training a large language model intelligent agent with dynamic reasoning ability in the comprehensive energy system field. Step S2 specifically comprises:
[0012] Step S21, basic capability pre-training; specifically, based on the pre-trained language model base, using the multi-source heterogeneous knowledge base of the integrated energy system, the mapping relationship between natural language instructions and structured parameters such as adjustable devices, scheduling time period, and scheduling target weight is established through supervised learning, and a large language model with semantic understanding capability of integrated energy system domain terminology and basic rules is obtained; and the semantic understanding capability of the large language model for the integrated energy system domain terminology and basic rules is further strengthened based on the rule logic coding system.
[0013] Step S22, complex reasoning capability reinforcement training; specifically, through a multi-task joint training mechanism, the large language model synchronously learns semantic parsing, time series trend reasoning, and dynamic constraint generation capability (specifically, the ability to dynamically adjust or generate new computable constraint conditions according to user semantics); at the same time, the large language model deduces mathematical constraint conditions matching the current scheduling scene based on real-time device state, user demand priority, and rule logic coding system, and calls the computable constraint condition set verification logic in the rule logic coding system to verify logical consistency, thereby improving the complex reasoning capability of the large language model, and finally enabling the large language model to realize end-to-end reasoning from fuzzy demand to accurate scheduling parameters.
[0014] Step S23, incremental learning interface deployment; specifically, the large language model dynamically accesses device change parameters, policy change texts, and user feedback information, and through a parameter efficient fine-tuning method, the large language model continuously iterates the cognitive ability, synchronously updates the rules and association relationships of the constraint conditions in the rule logic coding system (the rule refers to the mandatory limit of the constraint condition, such as the upper and lower limits of the device power; the association relationship refers to the logical relationship between the constraint conditions, such as whether they conflict with each other), and ensures the long-term adaptability of the reasoning capability of the large language model and the dynamic evolution of the system, thereby obtaining a large language model agent with dynamic reasoning capability in the integrated energy system field.
[0015] Step S3, using the large language model agent constructed in step S2 to parse the user input to obtain an integrated energy system scheduling demand, the integrated energy system scheduling demand including adjustable devices, scheduling time period, and scheduling target weight; and constructing an optimization scheduling scene based on the integrated energy system scheduling demand. Step S3 specifically comprises:
[0016] Based on the natural language understanding capability of the large language model, the user input text scheduling demand is converted into an integrated energy system scheduling demand, and if the user input demand does not mention the specified content, it is set as the default value when the large language model agent is constructed.
[0017] Based on the integrated energy system scheduling demand, an optimization scheduling scene is constructed, which is specifically represented as:
[0018] Adjustable device set is:
[0019]
[0020] wherein, is the set of all devices in the system, represents a device In real-time state at time t, state is 1, which is adjustable.
[0021] The scheduling period T is:
[0022]
[0023] wherein, is the scheduling time, is the scheduling start time, k is the scheduling number, and there are N scheduling times.
[0024] Scheduling target weight parameters are:
[0025]
[0026] wherein, is the energy cost weight, is the carbon emission cost weight.
[0027] Step S4, using the large language model agent to dynamically predict the carbon price and energy price in the scheduling period. Specifically:
[0028] Based on the large language model agent, using historical price data and user input information, generate carbon price prediction sequence and energy price prediction sequence in the scheduling period. User input information includes real-time policy text and supply and demand situation information; If the user does not provide specified information, generate the prediction sequence according to the default parameters of the large language model agent.
[0029] The carbon price prediction sequence is:
[0030]
[0031] wherein, is the time series trend analysis function, is the historical carbon price data, is the carbon price weight coefficient of user input information, is the semantic influence quantification function, is the user input information.
[0032] The energy price prediction sequence is:
[0033]
[0034] wherein m is the type of energy used by the device (including gas, steam, electricity), is the historical price data of the corresponding type of energy, is the energy price weight coefficient of the corresponding type of energy in the user input information.
[0035] Step S5, according to the optimized scheduling scenario and the predicted carbon price and energy price, a flexible optimization scheduling model containing an objective function and constraint conditions is established. Specifically:
[0036] Based on the parameters in the optimized scheduling scenario obtained in step S3 and the carbon price and energy price prediction sequence obtained in step S4, a flexible optimization scheduling model is constructed, which takes the weighted sum minimization of carbon cost and energy cost as the objective function, and takes the device output range constraint, device output climbing constraint, multi-energy flow balance constraint, and system topology structure constraint as the constraint condition. Solving the model can obtain the optimal output strategy of the adjustable device in the scheduling period. By analyzing the user input demand as the comprehensive energy system scheduling demand, the optimized scheduling scenario is constructed, and the carbon price and energy price prediction sequence are dynamically predicted. The objective function and constraint conditions established can be flexibly changed, thereby dynamically meeting the actual scheduling demand.
[0037] The objective function is:
[0038]
[0039] wherein, is the objective function value, is the output of device i in period, indicates that device i belongs to the adjustable device set , is the carbon emission coefficient of device i; the remaining parameters in the objective function are flexibly generated after analyzing the user input through the foregoing steps.
[0040] The constraint condition parameter set is included in the rule logic coding system when the large language model is constructed, and the flexible optimization scheduling model selects the constraint condition parameter subset corresponding to the adjustable device set therefrom.
[0041] Step S6, solving the flexible optimization scheduling model, and updating the large language model agent of the model in real time according to the user feedback and the scheduling result, and realizing the adaptive optimization of the scheduling strategy based on the large language model agent. Specifically:
[0042] Solving the flexible optimization scheduling model, obtaining the optimal output of each adjustable device at each scheduling time. Based on the further correction instructions of the user to the obtained scheduling strategy and the actual running data, the large language model agent analyzes the correction instructions and extracts the comprehensive energy system scheduling demand, dynamically updates the parameters of the flexible optimization scheduling model, and regenerates the scheduling strategy that meets the latest demand of the user.
[0043] At the same time, based on the correction instructions and the actual running data, the running deviation data is obtained, and the correction instructions and the running deviation data are input into the large language model agent, and the price prediction logic of step S4 and the scheduling demand analysis rule of step S3 are updated through the incremental learning mechanism, the multi-source heterogeneous knowledge base of the comprehensive energy system is optimized synchronously, and the subsequent scheduling strategy can adapt to external changes, and the effectiveness and robustness of the long-term scheduling effect are ensured.
[0044] Finally, the adaptive optimization of the scheduling strategy can be realized based on the large language model agent.
[0045] The application also provides a computer readable storage medium, which stores computer instructions for making a computer execute the flexible optimization scheduling method of the comprehensive energy system based on the large language model.
[0046] The beneficial effects of the application are:
[0047] The application realizes the dynamic coupling of scheduling demand and system state through the large language model agent, breaking through the limitations of traditional optimization methods relying on fixed models and static parameters. On the one hand, based on natural language interaction, the user can flexibly adjust the scheduling time scale, scheduling target weight and adjustable device range, and the optimized scene can be quickly generated without reconfiguring the model; on the other hand, through the incremental learning of the large language model, the scheduling strategy and the environmental changes are synchronized to evolve, avoiding the problem of invalid scheduling strategy. The method can be beneficial to improve the flexibility and adaptability of the optimization scheduling strategy of the comprehensive energy system, and provides technical support for building an efficient and low-carbon new energy system. BRIEF DESCRIPTION OF DRAWINGS
[0048] Figure 1 The main steps of the application are as follows:
[0049] Figure 2 The mechanism diagram of the flexible optimization scheduling of the comprehensive energy system based on the large language model agent is shown in the figure. DETAILED DESCRIPTION
[0050] The application will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, and only illustrate the basic structure of the application in a schematic manner, and therefore only show the components related to the application.
[0051] Example 1
[0052] As Figure 1 shown, the present application provides a comprehensive energy system flexible optimization scheduling method based on a large language model, and the main steps include:
[0053] In step S1:
[0054] Step S11, build a comprehensive energy system multi-source heterogeneous knowledge base; from the SCADA system, equipment manufacturer database and policy release platform of the comprehensive energy system, respectively collect equipment technical parameters (including rated power, efficiency), system topology structure (stored in the form of adjacency matrix), power market transaction rule documents and policy documents (PDF / text format, unstructured data, all of which are structured data), equipment and load historical data (CSV structured data), historical energy prices and carbon prices (CSV structured data); and unify the storage and association mapping of structured data and unstructured data, and establish a comprehensive energy system multi-source heterogeneous knowledge base supporting semantic retrieval.
[0055] Step S12, build a rule logic coding system; specifically: perform semantic analysis and entity relationship extraction on the unstructured data in the comprehensive energy system multi-source heterogeneous knowledge base, identify the operation rules, constraint conditions and policy provisions in the comprehensive energy system domain knowledge, and convert them into computer executable IF-THEN logic statements or JSON format files; through a dynamic mapping mechanism, the operation rules are bound with the real-time state variables of the system, a set of computable constraint conditions suitable for different scheduling scenarios are generated, and through an API interface, the set of computable constraint conditions are fused with the structured data in the comprehensive energy system multi-source heterogeneous knowledge base, and a rule logic coding system supporting multi-dimensional logical reasoning is constructed.
[0056] In step S2:
[0057] Step S21, basic capability pre-training; the specific method is: using a pre-training language model base with a Transformer architecture, using the comprehensive energy system multi-source heterogeneous knowledge base, establishing the mapping relationship between natural language instructions and structured parameters such as adjustable equipment, scheduling period, and scheduling target weight through supervised learning, and further strengthening the semantic understanding ability of the large language model for energy system domain terminology and basic rules based on the rule logic coding system.
[0058] Step S22, complex reasoning ability reinforcement training; the specific method is: through the multi-task joint training mechanism, the large language model synchronously learns the semantic analysis, time trend reasoning and dynamic constraint generation ability; at the same time, the large language model is based on real-time device state, user demand priority and rule logic coding system, deduces the mathematical constraint condition matched with the current scheduling scene, and calls the computable constraint condition set verification logic consistency, thereby improving the complex reasoning ability of the large language model, and finally making the large language model realize the end-to-end reasoning from fuzzy demand to accurate scheduling parameter.
[0059] Step S23, incremental learning interface deployment; the specific method is: the large language model dynamically accesses device change parameters, policy change text and user feedback information, realizes continuous iteration of the cognitive ability of the large language model through parameter efficient fine-tuning method, synchronously updates the rules and association relationships of constraint conditions in the rule logic coding system, ensures the long-term adaptability of the reasoning ability of the large language model and the dynamic evolution of the system, thereby obtaining the large language model intelligent agent with dynamic reasoning ability in the field of comprehensive energy system.
[0060] In step S3:
[0061] Based on the natural language understanding ability of the large language model, the user input text scheduling demand is converted into a comprehensive energy system scheduling demand, which specifically includes adjustable devices, scheduling time period and scheduling target weight; if the user input demand does not mention the specified content, it is set as the default value when the large language model intelligent agent is constructed. In this embodiment, when the user inputs "today from 8am to 6pm, need to prioritize reducing energy cost, 2 and 3 combined heat and power units and energy storage system can be dynamically adjusted, the rest of the devices still follow the day-ahead scheduling strategy", it is converted into a comprehensive energy system scheduling demand, and an optimal scheduling scene is constructed based on the comprehensive energy system scheduling demand, which can be specifically represented as:
[0062] Adjustable device set As follows:
[0063]
[0064] Among them, 2 and 3 represent the combined heat and power units and energy storage devices numbered 2 and 3 in the system topology library, respectively.
[0065] Scheduling time period As follows:
[0066]
[0067] Among them, is the scheduling time, is the scheduling start time, , k is the scheduling number, a total of N times; after analyzing the user input, Set to 8:00, Set to 30 minutes, the total number of scheduling is set to 20 times.
[0068] Build scheduling target weight coefficient As follows:
[0069]
[0070] In step S4:
[0071] Based on the large language model agent, the historical price data and the user input information are used to generate the carbon price prediction sequence and the energy price prediction sequence in the scheduling period; the user input information includes real-time policy text and supply and demand situation information; if the user does not provide specified information, the prediction sequence is generated according to the default parameters of the large language model agent. For the , since the user input does not mention real-time price policy and supply and demand situation information, the carbon price prediction sequence and the energy price prediction sequence are constructed according to the default parameters during the training of the large language model agent:
[0072]
[0073] Where, is the time series trend analysis function, is the historical carbon price data, is the carbon price weight coefficient of the user input information, is the semantic influence quantification function, is the user input information.
[0074]
[0075] Where, m is the type of energy used by the device (including gas, steam, electricity), is the historical price data of the corresponding type of energy, is the energy price weight coefficient of the corresponding type of the user input information.
[0076] In step S5:
[0077] According to the optimization scheduling scenario of step S3 and the carbon price and energy price predicted in step S4, a flexible optimization scheduling model containing an objective function and constraint conditions is established. Specifically, the flexible optimization scheduling model takes the weighted sum of carbon cost and energy cost minimization as the objective function, and takes the device output range constraint, device output ramp constraint, multi-energy flow balance constraint, and system topology structure constraint as the constraint conditions. After the user input demand is parsed into the comprehensive energy system scheduling demand, the optimization scheduling scenario is constructed, and the carbon price and energy price prediction sequence are dynamically predicted, so that the objective function and the constraint conditions can be flexibly changed to dynamically meet the actual scheduling demand. The objective function is specifically constructed as:
[0078]
[0079] wherein, is the objective function value, is the output of device i in the time period, ; represents that device i belongs to the adjustable device set ; is the carbon emission coefficient of device i.
[0080] The constraint condition parameter set is contained in the field knowledge and system data when the large language model is constructed, and the flexible optimization scheduling model can select the constraint condition parameter subset corresponding to the adjustable device set from the constraint condition parameter set.
[0081] In step S6: the flexible optimization scheduling model of step S5 is solved, and the large language model agent is updated in real time according to the user feedback and the scheduling result, and the adaptive optimization of the scheduling strategy is realized based on the large language model agent. As shown in Figure 2 , this step specifically includes:
[0082] solving the flexible optimization scheduling model to obtain the optimal output of the adjustable device at each scheduling time. Based on the further correction instructions of the user to the obtained scheduling strategy and the actual running data, the large language model agent is used to parse the correction instructions and extract the comprehensive energy system scheduling demand, dynamically update the parameters of the flexible optimization scheduling model, and regenerate the scheduling strategy that meets the latest demand of the user.
[0083] At the same time, the running deviation data is obtained based on the correction instructions and the actual running data, and the user correction instructions and the running deviation data are input into the large language model agent, and the price prediction logic of step S4 and the demand parsing rule of step S3 are updated through the incremental learning mechanism, the multi-source heterogeneous knowledge base of the comprehensive energy system is optimized synchronously, and the subsequent scheduling strategy can adapt to external changes, and the effectiveness and robustness of the long-term scheduling effect are ensured.
[0084] Embodiment 2
[0085] The application further provides a computer readable storage medium, which stores computer instructions for making a computer execute the flexible optimization scheduling method of the integrated energy system based on a large language model.
[0086] It should be noted that the above description is only part of the embodiments of the present application, and equivalent changes made by the system described in the present application are included in the protection scope of the present application. Those skilled in the art can make similar substitutions to the described specific examples, as long as they do not deviate from the structure of the present application or exceed the scope defined by the present claims, and they belong to the protection scope of the present application.
Claims
1. A flexible optimization scheduling method for integrated energy systems based on a large language model, characterized in that, Includes the following steps: Step S1: Construct a multi-source heterogeneous knowledge base and rule logic coding system for the integrated energy system; Step S2: Based on the multi-source heterogeneous knowledge base and rule logic encoding system of the integrated energy system, train a large language model intelligent agent with dynamic reasoning ability in the field of integrated energy system; Step S3: Use the large language model agent to parse the user input to obtain the integrated energy system scheduling requirements, which include adjustable equipment, scheduling time periods and scheduling target weights. And based on the scheduling needs of the integrated energy system, an optimized scheduling scenario is constructed; Step S4: Dynamically predict carbon price and energy price during the scheduling period using the large language model agent; Step S5: Based on the optimized scheduling scenario described in Step S3 and the carbon price and energy price predicted in Step S4, establish a flexible optimized scheduling model that includes an objective function and constraints. Step S6: Solve the flexible optimization scheduling model described in step S5, and update the large language model agent in real time based on user feedback and scheduling results, and realize adaptive optimization of the scheduling strategy based on the large language model agent. In step S5, the objective function is to minimize the weighted sum of carbon cost and energy cost, and the constraints are equipment output range constraint, equipment output ramping constraint, multi-energy flow balance constraint, and system topology constraint. The objective function is: ； in, Let k be the objective function value, and k be the number of scheduling operations, with a total of N scheduling operations. For scheduling time, For device i in Efforts during a specific time period This indicates that device i belongs to the set of adjustable devices. , Let be the carbon emission coefficient of device i; As a weighted average of energy costs, As a weight for carbon emission costs; This is a carbon price prediction sequence.
2. The flexible optimization scheduling method for a comprehensive energy system based on a large language model according to claim 1, characterized in that, Step S1 specifically includes the following steps: Step S11: Construct a multi-source heterogeneous knowledge base for the integrated energy system. Specifically, this involves integrating structured and unstructured data, and simultaneously storing and mapping the structured and unstructured data in a unified manner to establish a multi-source heterogeneous knowledge base for the integrated energy system. This multi-source heterogeneous knowledge base supports semantic retrieval. The structured data includes technical parameters of integrated energy system equipment, system topology, historical equipment and load data, historical energy prices, and carbon prices. The unstructured data includes trading mechanism rules and policy texts. Step S12 involves constructing a rule logic encoding system. Specifically, this involves: performing semantic parsing and entity relation extraction on unstructured data in the multi-source heterogeneous knowledge base of the integrated energy system; identifying the operating rules, constraints, and policy clauses in the knowledge of the integrated energy system domain; and converting the operating rules, constraints, and policy clauses into computer-executable formal logical expressions. A dynamic mapping mechanism is used to bind the operating rules to real-time system state variables, generating a set of computable constraints adapted to different scheduling scenarios. This set of computable constraints is then integrated with the structured data in the multi-source heterogeneous knowledge base of the integrated energy system to obtain the rule logic encoding system. The rule logic encoding system supports multi-dimensional logical reasoning.
3. The flexible optimization scheduling method for a comprehensive energy system based on a large language model according to claim 2, characterized in that, In step S2: Step S21, basic capability pre-training; specifically: based on the pre-trained language model base, using the multi-source heterogeneous knowledge base of the integrated energy system, a mapping relationship between natural language instructions and adjustable equipment, scheduling time periods, and scheduling target weights is established through supervised learning to obtain a large language model with semantic understanding capabilities of integrated energy system domain terms and basic rules; and based on the rule logic encoding system, the semantic understanding capability of the large language model of integrated energy system domain terms and basic rules is further enhanced. Step S22, complex reasoning ability enhancement training; specifically, through a multi-task joint training mechanism, the large language model synchronously learns semantic parsing, temporal trend reasoning and dynamic constraint generation capabilities; at the same time, the large language model derives mathematical constraints that match the current scheduling scenario based on real-time device status, user demand priority and rule logic encoding system, and calls the computable constraint set to verify logical consistency, thereby improving the complex reasoning ability of the large language model; Step S23, deploy the incremental learning interface; Specifically, the large language model dynamically accesses device parameter changes, policy change texts, and user feedback information. Through efficient parameter fine-tuning methods, the cognitive ability of the large language model is continuously iterated. The rules and relationships of constraints in the rule logic coding system are updated synchronously to ensure the long-term adaptability of the large language model's reasoning ability to the dynamic evolution of the system, thereby obtaining a large language model intelligent agent with dynamic reasoning ability in the field of integrated energy systems.
4. The flexible optimization scheduling method for a comprehensive energy system based on a large language model according to claim 1, characterized in that, The specific steps of step S3 are as follows: Based on the natural language understanding capabilities of the large language model, the user's input text scheduling requirements are transformed into integrated energy system scheduling requirements. If the user's input does not mention the specified content, it is set to the default value when the large language model agent is constructed. An optimized scheduling scenario is constructed based on the scheduling requirements of the integrated energy system, specifically represented as follows: Adjustable device collection for: ； in, For all devices in the system, Indicates equipment In the real-time state at time t, a state of 1 indicates that it is adjustable; The scheduling period T is: ； in, For scheduling time, The scheduling start time, k represents the number of scheduling operations, and a total of N scheduling operations are performed. Scheduling target weight parameters for: ； in, As a weighted average of energy costs, Carbon emission cost weighting.
5. The flexible optimization scheduling method for a comprehensive energy system based on a large language model according to claim 4, characterized in that, The specific steps of S4 are as follows: Based on a large language model agent, the system generates carbon price and energy price prediction sequences for the specified time period using historical price data and user input information. The user input information includes real-time policy text and supply and demand information. If the user does not provide specified information, the prediction sequence is generated according to the default parameters of the large language model agent.
6. The flexible optimization scheduling method for a comprehensive energy system based on a large language model according to claim 5, characterized in that, The carbon price prediction sequence for: ； in, This is a time-series trend analysis function. Historical carbon price data, The carbon price weighting coefficient for user-input information. For semantic influence quantization function, Enter information for the user; The energy price forecast series for: ； Where m represents the type of energy used by the equipment. For the corresponding type of historical energy price data, Provide the corresponding energy price weighting coefficient for the information entered by the user.
7. The flexible optimization scheduling method for a comprehensive energy system based on a large language model according to claim 1, characterized in that, In step S6: The large language model agent is updated in real time based on user feedback and scheduling results. The specific method is as follows: Based on the user's further correction instructions for the obtained scheduling strategy and actual operating data, the large language model agent parses the correction instructions and extracts the scheduling requirements of the integrated energy system. The scheduling model parameters are dynamically updated and flexibly optimized to regenerate a scheduling strategy that meets the user's latest needs. At the same time, based on the correction instructions and actual operating data, the operating deviation data is obtained and input into the large language model agent. The price prediction logic in step S4 and the scheduling requirement parsing rules in step S3 are updated through an incremental learning mechanism. The multi-source heterogeneous knowledge base of the integrated energy system is optimized in parallel to ensure that the subsequent scheduling strategy can adapt to external changes.
8. A computer-readable storage medium storing computer instructions thereon, characterized in that, The computer instructions are used to cause the computer to perform the steps of the method as described in any one of claims 1-7.