Control method and system for comprehensive energy supply device of intelligent calculation center

Abstract

The invention relates to the technical field of computer systems based on specific calculation models, and discloses a control method and system for an intelligent calculation center comprehensive energy supply device, and the method comprises the steps: collecting the operation data of energy supply equipment and an environment sensor in real time through an SCADA system, and carrying out the preprocessing; based on the collected data and simulation data generated by a simulation environment, performing offline mixed training on the reinforcement learning model to generate an energy scheduling strategy; inputting a real-time state into the trained reinforcement learning model to generate a preliminary scheduling instruction, performing security verification and interpretability analysis by using a large language model, and optimizing a strategy; and fusing the preliminary scheduling instruction with the suggestion of the large language model, generating a final scheduling command through the energy router control unit, and issuing the final scheduling command to the energy supply equipment for execution. The problems that in the prior art, black box decision making and simulation are not accurate, and experience is difficult to solidify are solved, and the purposes of decision making transparency, simulation high fidelity and experience structuring are achieved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computer systems based on specific computing models, and in particular to reinforcement learning. BACKGROUND

[0002] With the continuous expansion of the scale of the intelligent computing center, it is necessary to integrate various forms of energy supply to meet the diversified demand for power and refrigeration. The existing intelligent management and control platform mainly monitors the running state and energy consumption of various energy equipment in real time through the SCADA system, predicts the energy supply and demand, and intelligently schedules, and then closes the loop to issue control instructions to each physical device to ensure the all-weather power supply and cooling demand of the intelligent computing center, thereby optimizing energy distribution, improving operation efficiency and energy utilization rate. Traditional fixed value control or optimization methods based on theoretical physical models are difficult to quickly adapt to real-time changes in load and energy supply conditions, limiting the play of system intelligence and adaptive scheduling capabilities. With the advancement of intelligent energy supply systems, existing research and inventions have shown that applying intelligent algorithms such as reinforcement learning to integrated energy systems can achieve coordinated scheduling and unified planning of multiple energies, effectively improving energy utilization efficiency and reducing operating costs. However, data-driven methods often lack transparent explanations of complex decision-making processes, and their reliance on simulation environments may deviate from the actual physical system, affecting the real effectiveness of the control strategy.

[0003] For example, the Chinese patent with publication number CN118519340B discloses an intelligent heating regulation method based on deep reinforcement learning and adaptive control, which provides the following technical solution. The present application relates to the technical field of heating pipe network regulation, and discloses an intelligent heating regulation method based on deep reinforcement learning and adaptive control, comprising the following steps: S1: collecting outdoor environment data and heat load data of each heat station, and performing data preprocessing; S2: training a deep reinforcement learning model based on the outdoor environment data and heat load data; S3: developing a heating strategy based on the deep reinforcement learning model; S4: executing the heating strategy through an adaptive control algorithm. The present application improves energy utilization efficiency and reduces unnecessary energy loss; enhances the stability and reliability of the entire heating system, ensures heating quality, and optimizes energy distribution. However, the above-mentioned intelligent heating regulation method based on deep reinforcement learning and adaptive control relies on historical heat load mean to generate actions, which is difficult to adapt to abnormal working conditions beyond the historical range; the heat loss coefficient is fixed and cannot dynamically respond to changes in the pipe network, and the reinforcement learning decision lacks interpretability, with low operation transparency. SUMMARY

[0004] The present application solves the problems of black box decision, simulation misalignment and difficulty in solidifying experience in the prior art, and proposes a control method and system for a comprehensive energy supply device of an intelligent computing center, achieving the purposes of decision transparency, high-fidelity simulation, and experience structuring.

[0005] Further, the present application significantly improves the energy utilization efficiency of the intelligent computing center by fusing reinforcement learning with a large language model, reduces operating costs, provides interpretable decision-making to enhance transparency, and ensures system adaptive scheduling and safe and reliable operation, achieving high efficiency and energy saving.

[0006] To achieve the above-mentioned purposes, the present application adopts the following technical solutions: A control method for a comprehensive energy supply device of an intelligent computing center, comprising: Collecting and preprocessing the operation data of the energy supply equipment and environmental sensors in real time through the SCADA system; Based on the collected data and simulation data generated by the simulation environment, the reinforcement learning model is trained offline to generate an energy scheduling strategy; Inputting the real-time state into the trained reinforcement learning model to generate preliminary scheduling instructions, and using a large language model for safety verification and interpretable analysis, and optimizing the strategy; Fusing the preliminary scheduling instructions and the suggestions of the large language model, generating the final scheduling command through the energy router control unit, and issuing it to the energy supply equipment for execution.

[0007] Combining real and simulation data to optimize model training enhances the generalization ability and robustness of the strategy, reduces actual system risks through offline safety environment, improves scheduling accuracy and the ability to adapt to complex and variable scenarios, fuses intelligent algorithms and natural language processing to ensure decision transparency and understandability, reduces operational risks through safety verification and optimizes the strategy, significantly improves operator trust and overall system performance.

[0008] A control system for a comprehensive energy supply device of an intelligent computing center, comprising: Energy supply equipment, including power supply module and cooling module; SCADA monitoring unit, collecting equipment operation state and multi-modal data in real time, connected with the energy supply equipment; Reinforcement learning control unit, training model based on cache data and outputting scheduling strategy; Large language model unit, evaluating the safety of the strategy and providing interpretable analysis and optimization suggestions, connected with the reinforcement learning control unit and the SCADA monitoring unit; Energy router control unit, executing dynamic scheduling commands, and connected with the large language model unit and the energy supply equipment.

[0009] A complete closed-loop control system architecture is provided, which realizes efficient data flow and intelligent decision-making through modular design, enhances system integration and scalability, and ensures adaptive optimization and continuous improvement of energy supply.

[0010] As a preferred, the reinforcement learning model generating preliminary scheduling instructions specifically includes inputting real-time state into the reinforcement learning model, the model considering real-time constraints, outputting corresponding actions after forward calculation, and finally outputting the energy scheduling decision-making instruction at the current time predicted by the reinforcement learning model; the real-time state includes data from each energy device and the intelligent computing center at the current time, specifically including the output situation and available capacity of the energy device and the power load and cooling load demand of the intelligent computing center.

[0011] The feasibility and efficiency of the scheduling instruction under real-time constraints are ensured, key operating parameters are covered through comprehensive state input, optimized actions are quickly generated, and problems such as device overload or resource shortage are avoided, thereby improving system stability and reliability.

[0012] As a preferred, the security verification and explainability analysis using a large language model includes judging whether the instruction conforms to the operating constraints and labeling potential risks according to the input strategy and system information, then converting the complex decision-making logic inside the model into natural language description, pointing out the key factors affecting the decision, and extracting implicit experience rules in system operation; the strategy optimization includes providing strategy fine-tuning suggestions based on identified model bias and business objectives, and generating contextual explanations and counterfactual analysis of strategy effectiveness; the system information includes current system state and historical scheduling data as well as system operation logs.

[0013] The transparency and traceability of the decision-making process are improved, helping operators understand complex strategies and respond quickly to risks, adapting strategies through experience rule extraction and fine-tuning optimization, reducing actual deployment errors, and enhancing system intelligence level.

[0014] As a preferred, the fusion of preliminary scheduling instructions and large language model suggestions includes combining the scheduling strategy given by the reinforcement learning model with the suggestions of the large language model by the energy router control unit, using weighted average and heuristic rules and suboptimal decision tree strategy, combining reinforcement learning output and experience rules to form the final command; the issuing to energy supply devices for execution includes the energy router issuing commands through the control bus or communication network to each device to implement dynamic allocation of power load and cooling load.

[0015] Intelligent fusion of multi-source decision-making is realized, efficient communication network is ensured to ensure fast execution of commands, energy distribution is optimized, and peak-valley filling scheduling is realized, significantly reducing operating costs and improving sustainability.

[0016] Preferably, the simulation data is generated by a simulation environment, simulating the operating state and energy output and consumption of several energy equipment and cold source equipment under different loads and environmental conditions, and the input data includes a multi-dimensional state vector and possible action scheme and corresponding feedback; the offline hybrid training includes offline learning of the reinforcement learning algorithm based on the input sample, the agent selects a scheduling action in the virtual environment through repeated trial and error, obtains a reward from the environment, and optimizes the policy network accordingly, and finally outputs the trained reinforcement learning scheduling model, the updated model parameters and the performance indicators.

[0017] By covering various scenarios through a high-fidelity simulation environment, the model training effect is enhanced, the trial and error cost of the real system is avoided, the policy robustness is improved through offline learning iteration, and the scheduling model is ensured to be efficient and reliable in actual deployment.

[0018] Preferably, the operating data includes the operating state of the multi-energy equipment and load in the intelligent algorithm center, specifically including equipment state information and operation log and real-time sensing data and visual perception data, and the preprocessing includes time synchronization of data from different sources, analysis of the transmitted data, and pushing them to real-time data cache or database.

[0019] The consistency and integrity of the data are ensured, laying a solid foundation for analysis and training, reducing the influence of noise and missing values through preprocessing, improving the overall data-driven capability of the system, and supporting more accurate energy management decisions.

[0020] Preferably, the power supply module includes a gas turbine power generation device and an electrochemical energy storage device and an auxiliary power supply, the gas turbine power generation device generates electric energy using a natural gas power device, and at the same time outputs steam to a steam-type refrigerator, the auxiliary power supply includes a power grid and a photovoltaic power generation device, and the electrochemical energy storage device stores excess electric energy and supplements the power supply during peak load.

[0021] Efficient collaborative use of multi-source energy is achieved, system reliability and flexibility are improved through integration of energy storage and renewable energy, dependence on external power grid is reduced, energy cost is optimized, and sustainability is enhanced.

[0022] Preferably, the SCADA data monitoring system collects the operating state and multi-modal data of each energy supply equipment in real time and writes them into the real operating data cache area; the simulation data cache area stores historical energy supply and demand simulation data generated by the digital twin simulation environment, and the reinforcement learning control unit obtains data from the real operating data cache area and the simulation data cache area, and performs hybrid training on the reinforcement learning model, thereby generating preliminary scheduling instructions.

[0023] Combined with real and simulated data reinforcement training process, improve model accuracy and adaptability, through digital twin environment calibration to reduce policy bias, ensure that the scheduling instruction remains efficient and stable under changing conditions.

[0024] As preferred, the cold supply module includes the steam-type refrigerator and the electrically driven centrifugal refrigerator, the steam-type refrigerator drives the refrigeration cycle by means of high-temperature steam generated by the gas turbine power generation device or the internal combustion engine, and the electrically driven centrifugal refrigerator refrigerates with system power, and the cold quantity output by the cold supply module is delivered to the intelligent computing center through the closed circulating cooling water system including the closed cooling water cooler and the cold water storage tank.

[0025] The cold energy production and delivery efficiency is optimized, the cold quantity loss and energy consumption are reduced through the closed system, the high-intensity cooling demand of the intelligent computing center is met, the overall energy utilization rate is improved, and the operation cost is reduced.

[0026] Compared with the prior art, the beneficial effects of the present application are.

[0027] 1. The present application solves the "black box decision" problem of traditional data-driven methods by deeply evaluating and explaining reinforcement learning strategies through a large language model unit. The LLM unit analyzes the context of the scheduling instruction, converts complex decision logic into natural language description, and extracts structured experience rules. This significantly improves the operator's understanding and trust of the system behavior, reduces the risk of human error, and optimizes the adaptability of the strategy through fine-tuning suggestions. In practical applications, this mechanism makes the scheduling process traceable and auditable, enhancing the reliability and safety of the system under varying loads.

[0028] 2. The present application generates high-fidelity data using a digital twin simulation environment, and combines real operation data collected by SCADA for offline hybrid training, significantly improving the generalization ability and robustness of the reinforcement learning model. Specifically, the simulation environment simulates diverse scenarios, inputs multi-dimensional state vectors and action schemes, and the reinforcement learning algorithm optimizes the strategy network through repeated trial and error, and outputs updated model parameters and performance indicators. The LLM unit further calibrates simulation bias to ensure that the training environment is close to the actual physical system. This reduces the risk of strategy misalignment in actual deployment and improves the system's adaptability to complex and variable conditions.

[0029] 3.The application realizes intelligent optimization distribution of multi-source energy through a dynamic scheduling mechanism combining reinforcement learning and large language model, significantly improves energy efficiency and reduces operating costs. The energy router control unit combines preliminary scheduling instructions and LLM recommendations to generate final commands, which are issued to energy supply equipment to perform peak-valley filling valley scheduling. At the same time, the power supply module and the cooling module work together through a closed cycle system to maximize the utilization rate of renewable energy. In practical applications, the system adaptively adjusts energy output, reducing dependence on external power grids and fossil fuel consumption. Overall, the application improves the energy utilization rate of the intelligent computing center, reduces the average annual operating cost, and ensures efficient and energy-saving operation of the system under safety constraints. BRIEF DESCRIPTION OF DRAWINGS

[0030] Figure 1 The control method and system of the application are shown in the overall flowchart of the intelligent computing center comprehensive energy supply device.

[0031] Figure 2 The control system module architecture diagram of the control method and system of the application is shown in the control system module architecture diagram of the intelligent computing center comprehensive energy supply device.

[0032] Figure 3 The intelligent computing center comprehensive energy supply system structure diagram of the control method and system of the application is shown in the intelligent computing center comprehensive energy supply system structure diagram of the intelligent computing center comprehensive energy supply device. DETAILED DESCRIPTION

[0033] To make the purpose, technical solution and advantages of the present disclosure clearer, the following will further describe the embodiments of the present disclosure in combination with the drawings. The proportions of the components are not drawn according to the true proportions, and the proportions and sizes shown in the drawings should not limit the essential technical solutions of the application. These embodiments do not describe all the details, and the application is not limited to the specific embodiments described.

[0034] Referring to Figures 1-3 The control method of the intelligent computing center comprehensive energy supply device is shown in the control method of the intelligent computing center comprehensive energy supply device, which includes: Real-time collection of operation data of energy supply equipment and environmental sensors through the SCADA system, and preprocessing; Based on the collected data and simulation data generated by the simulation environment, the reinforcement learning model is trained offline to generate an energy scheduling strategy; Input the real-time state into the trained reinforcement learning model to generate preliminary scheduling instructions, and use the large language model for safety verification and explainability analysis, and optimize the strategy; Combine the preliminary scheduling instructions with the suggestions of the large language model, and generate the final scheduling command through the energy router control unit, and issue it to the energy supply equipment for execution.

[0035] A control system of a comprehensive energy supply device of an intelligent center, comprising: The energy supply device comprises a power supply module and a cooling supply module. A SCADA monitoring unit is connected with the energy supply device and is configured to collect real-time operation state and multi-modal data of the device. A reinforcement learning control unit is configured to train a model based on buffer data and output a scheduling strategy. A large language model unit is connected with the reinforcement learning control unit and the SCADA monitoring unit, and is configured to evaluate the safety of the strategy and provide explainable analysis and optimization suggestions. An energy router control unit is connected with the large language model unit and the energy supply device, and is configured to execute a dynamic scheduling command.

[0036] The present application provides an intelligent center comprehensive energy supply system based on reinforcement learning and large language model and a control method thereof. The system structure comprises a plurality of module units for realizing multi-energy supply and intelligent scheduling. The system mainly comprises: a multi-source power supply device (such as a photovoltaic power generation device, an electrochemical energy storage device, a gas turbine and a power grid), a variety of refrigeration equipment (such as a natural cold source device, a centrifugal water chiller, an absorption water chiller and a cold storage tank), a SCADA monitoring unit, a reinforcement learning control unit, an energy router and an LLM unit.

[0037] In the cooperative optimization control of the device operation, the SCADA monitoring unit is responsible for collecting the operation state and energy consumption data of each device. The reinforcement learning control unit trains a reinforcement learning model based on the collected real-time data and generates an energy scheduling strategy. The LLM unit evaluates and explains the strategy output by the reinforcement learning unit and makes corresponding fine-tuning. The unit is responsible for receiving the final output scheduling instruction of the energy router, dynamically controlling the output of each energy supply device and refrigeration equipment, realizing the optimized allocation of power and cold load and accessing the energy internet.

[0038] The LLM unit plays a core role in intelligent analysis and optimization in the present application, and its main functions further include: Enhanced decision evaluation and explanation: This unit not only performs interpretable analysis on the scheduling policies executed by the reinforcement learning control unit, extracts operational experience rules, and generates policy fine-tuning suggestions, but also generates context-aware, multi-dimensional explanations for scheduling decisions. This includes translating the complex internal decision logic of reinforcement learning models (such as policy values and action probability distributions) into natural language descriptions that are easy for humans to understand, identifying key factors that influence decisions (similar to feature importance analysis), and even providing counterfactual explanations (e.g., "If the photovoltaic output had not decreased, the electrochemical energy storage discharge start time would have been delayed X minutes"), thereby significantly improving the understanding and trust of system behavior by operators. This deep explanation capability helps to bridge the gap between the "black box" nature of reinforcement learning and the actual operational needs of operators.

[0039] Experience rule extraction and application of the system: The LLM unit systematically extracts and formalizes operational experience rules by analyzing historical scheduling data, SCADA logs, operator input text records (such as shift handover records, fault descriptions), and potentially other multi-modal data (such as equipment inspection pictures, maintenance reports). These rules are structured and stored (e.g., building a knowledge graph), and actively used by the LLM to evaluate the rationality of newly generated reinforcement learning strategies, or to provide decision support and warnings to operators.

[0040] Simulation environment evaluation and calibration based on multi-modal data: The LLM unit uses multi-modal data inputs, including text (such as equipment maintenance logs, operation manuals, weather reports), images (such as equipment thermal imaging photos, abnormal working condition pictures captured by surveillance cameras), and sensor logs (which may include special sensor data such as vibration and sound in addition to standard SCADA data), to evaluate and calibrate the accuracy of the simulation environment used to train the reinforcement learning model. The unit uses advanced multi-modal data fusion techniques (such as attention mechanism-based fusion models or joint embedding methods) to integrate information from heterogeneous data sources, forming a more comprehensive and accurate portrait of the physical system's state and dynamic behavior. By comparing this fused real-world representation with the output of the simulation environment, the LLM can identify deviations between the two, such as the simulation model's failure to accurately reflect the energy efficiency degradation of a specific device under certain working conditions (as determined by maintenance logs and thermal imaging images), or its failure to capture the combined impact of extreme weather on renewable energy output. Based on these identified deviations, the LLM can suggest or automatically perform adjustments to the simulation environment parameters or improvements to the model structure, thereby significantly improving the fidelity of the simulation environment and ensuring that the reinforcement learning model is trained in a simulation environment that is closer to the real world, reducing the gap between the strategy and the simulation environment when deployed in practice.

[0041] Fine-tuning of reinforcement learning models based on multi-modal data and empirical rules: The LLM unit utilizes its deep analysis results of multi-modal data and extracted operational empirical rules to fine-tune and optimize the decision-making model of the reinforcement learning control unit in a refined and multi-path manner. Specific fine-tuning mechanisms include: LLM-guided reward function shaping: The high-level semantic understanding obtained from multi-modal data analysis (e.g., "when monitoring data of device A is abnormal and maintenance records show that it will be overhauled soon, device B should be enabled first") or extracted empirical rules (e.g., "the single start-up running time of device C should not be less than 30 minutes to avoid excessive wear and tear") are converted into dynamic adjustments or additional items of the reward function of the reinforcement learning agent, thereby guiding the agent to learn more robust, efficient or compliant with actual operation constraints strategies.

[0042] LLM-driven strategy criticism and improvement: LLM analyzes the behavior trajectory of the reinforcement learning agent, compares it with the ideal behavior pattern inferred based on multi-modal data or the extracted empirical rules, identifies potential suboptimal behavior or risks, and provides feedback to the reinforcement learning model to adjust its strategy network, such as suggesting alternative action sequences or correcting unreasonable exploration directions.

[0043] LLM-assisted hyperparameter adjustment and exploration strategy optimization: Based on monitoring of the reinforcement learning training process (e.g., learning curve, convergence speed, stability) and evaluation of the complexity of the current operating environment (through multi-modal data analysis), LLM can make optimization suggestions for the hyperparameters (such as learning rate, discount factor) of the reinforcement learning algorithm or the exploration strategy.

[0044] The above-mentioned modules cooperate with each other to form the overall structure of the system of the present application. The control method of the present application adopts a closed-loop optimization strategy. In the offline stage, the system uses the historical operation data generated by the high-fidelity simulation model continuously calibrated by the LLM in combination with the real-time collected SCADA data to build a reinforcement learning training environment and train the scheduling model; in the online stage, the reinforcement learning model performs reasoning based on real-time data, outputs multi-energy scheduling instructions, which are executed by the energy router in real time, to realize the adaptive allocation of the load. Compared with the traditional control scheme, the system can autonomously optimize the scheduling strategy through reinforcement learning, can respond to load and energy supply changes in real time, can improve energy utilization efficiency and reduce operating costs. The energy router dynamically switches the outputs of various energies to realize peak-valley filling and valley scheduling, significantly improving the overall energy efficiency of the system. In addition, the fusion of the deep analysis, interpretation and optimization functions of the LLM makes the scheduling decision-making process more transparent, and can realize continuous iterative improvement of the control strategy through means such as strategy interpretation, experience extraction, simulation environment calibration and reinforcement learning model fine-tuning, significantly enhancing the autonomous learning and environmental adaptation capabilities of the system. The system and method described in the present application fully utilize the synergistic advantages of multiple energies to realize intelligent, efficient and adaptive operation of comprehensive energy. The reinforcement learning algorithm combined with real-time SCADA data is used to perform closed-loop optimization control on multi-source energy, and the LLM is introduced for deep empowerment, which can effectively improve resource utilization and reduce network construction and operation costs.

[0045] As shown in an embodiment, Figure 1 As shown in an embodiment, Figure 1 is the overall flowchart of the control method and system of the comprehensive energy supply device of the intelligent computing center. First, the operation data of the energy supply equipment (including the gas turbine power generation device of the power supply module, the electrochemical energy storage device, the auxiliary power supply, and the steam type refrigeration machine, the electrically driven centrifugal refrigeration machine of the cooling supply module) and the environmental sensors in the intelligent computing center are collected in real time through the SCADA system, specifically covering equipment state information, operation logs, real-time sensing data and visual perception data; then, preprocessing is performed, including time synchronization of data from different sources, analysis of transmission data and pushing to real-time data cache or database.

[0046] Then, based on the real operation data collected by the SCADA and the simulation data generated by the simulation environment (the simulation data simulates the operation state, energy output and consumption of the energy equipment and the cold source equipment under different loads and environmental conditions, and the input contains a multi-dimensional state vector, an action scheme and corresponding feedback), the reinforcement learning model is trained offline: the reinforcement learning algorithm repeatedly tries and errors based on the input samples, the agent selects scheduling actions in the virtual environment and obtains reward feedback, according to which the policy network is optimized, and finally the trained reinforcement learning scheduling model and the updated model parameters and performance indicators are output.

[0047] Next, the real-time state (including the output of the energy equipment, available capacity, and the power load and cooling load demand of the intelligent computing center) is input into the trained reinforcement learning model, and the model outputs preliminary scheduling instructions after forward calculation considering real-time constraints. At the same time, the large language model is used to verify the safety of the instructions and analyze the interpretability: according to the input strategy, the current system state, historical scheduling data and operation logs, it is judged whether the instructions meet the operation constraints and potential risks are marked, the model decision logic is converted into natural language description and the key influencing factors are pointed out, and the implicit experience rules are extracted; based on the identified model bias and business objectives, policy tuning suggestions are provided, and context explanations and counterfactual analysis of policy effects are generated.

[0048] Finally, the energy router control unit adopts weighted average, heuristic rules and suboptimal decision tree strategy to integrate the preliminary scheduling instructions output by the reinforcement learning model and the optimization suggestions of the large language model to form the final scheduling command; the command is issued to the energy supply equipment through the control bus or communication network to perform dynamic allocation of power load and cooling load.

[0049] In another embodiment, the method flow of the present application includes the following five main steps: The first step: data collection input: raw operation data of each energy supply equipment and environmental sensors, including equipment state information, operation logs, real-time sensor data and visual perception data, etc. For example, the output current and voltage of photovoltaic modules; the state of charge (SOC) of batteries; the output power of gas turbines; the refrigerating capacity of water chillers; environmental temperature and humidity and equipment thermal imaging pictures; and the power and cooling load data inside the intelligent computing center, etc. These data types are diverse, including time series numerical data (such as temperature, power, and electricity), state flags (such as equipment on-off signals), text log entries (such as alarm logs and operation records), and image data (such as thermal imaging photos).

[0050] Output: cleaned and formatted structured monitoring data stream or database records. For example, raw sensor data is organized into time series data tables with fields including timestamp, device ID, measurement type and value, etc.; log texts are stored in databases and indexed by time; image data is associated with corresponding time points and devices. These output data types may be CSV tables, JSON records or entries in time series databases, which are convenient for subsequent analysis and training use.

[0051] Core logic: Use SCADA (Supervisory Control and Data Acquisition) system to continuously scan and record the running status of multiple energy devices and loads in the intelligence center. The system collects data in real time through PLC or sensor network and performs preliminary verification and preprocessing (such as denoising, missing value completion, unit conversion). The data collection module synchronizes the data from different sources in time, analyzes the transmitted data, and pushes it to the real-time data cache or database. The collection content covers power grid input, power output, cold source output, and environmental feedback, etc., to ensure the coverage of global information required for comprehensive energy such as electricity and cold energy.

[0052] The present embodiment can refer to Table 1 below.

[0053] Table 1: Second step: offline training input: data collection results from the first step (real running data), and simulation data based on system model. The simulation data is generated by a simulation environment (or a high-fidelity simulator) to simulate the running status, energy output and consumption of multiple energy devices (photovoltaic, battery, gas turbine, etc.) and cold source devices under different loads and environmental conditions. The input data includes multi-dimensional state vectors (such as the output power of each energy device, environmental temperature load, etc.), possible action plans (such as the start and stop of each device or the output ratio), and corresponding feedback (simulated energy consumption, cost or comfort, etc.). These data types are usually in the form of matrices or tensors, describing a large number of samples of "<state, action, reward, next state>" sequences.

[0054] Output: trained reinforcement learning scheduling model and updated model parameters (policy network, value network, etc.). The output is in the form of a trained neural network parameter file or a policy function, so that the subsequent online phase can quickly infer control instructions based on the model. In addition, the output also includes performance indicators (such as cumulative income, loss curve, etc.) during the training process, and possible suggestions for improving the strategy.

[0055] Core logic: build reinforcement learning training environment and integrate real data and simulation data. The simulation environment generates a large number of training samples based on the physical model of the energy system and historical operation mode, covering different load fluctuations and weather change scenarios. Reinforcement learning algorithms (such as DDPG, SAC or PPO, etc.) learn offline based on these samples: through repeated trial and error, the agent selects scheduling actions (such as adjusting the output of the generator, charging and discharging strategy, etc.) in the virtual environment, obtains rewards (such as energy efficiency improvement or cost reduction) from the environment, and optimizes the policy network accordingly. During the training process, the simulation environment parameters can be corrected according to the real feedback and training results to ensure the authenticity of the model.

[0056] The training data samples of the present embodiment are shown as follows (JSON format):

[0057] Third step: Online decision input: Real-time system state data at the moment of operation, mainly from the SCADA system in the first step. Specifically, it includes the output and available capacity of each energy device (photovoltaic, energy storage, gas turbine, chiller, etc.) at the current moment; the power load and cooling load demand of the intelligent algorithm center; environmental data (such as outdoor temperature, humidity, and light intensity); and operation constraints. These inputs are usually in the form of real-time data streams into the trained reinforcement learning model, with data types mainly being numerical vectors (state vectors) or complex structures including state-environment descriptions.

[0058] Output: The energy scheduling decision instructions predicted by the reinforcement learning model at the current moment. For example, assign output power or control signals to each device: set the output power of photovoltaic generators, charge and discharge rate of batteries, start and stop or power setting of gas turbines, and on-off state of cooling devices. The output can be a vector or structured command containing the scheduling scheme for each device. Usually, it is an action vector, where each element corresponds to the control quantity (numerical value, percentage, or on-off instruction) of a device or load.

[0059] Core logic: Use the trained reinforcement learning strategy network from the second step to infer the optimal scheduling strategy in real time. Input the current state into the reinforcement learning model, and the model outputs the corresponding action after forward calculation. For example, the neural network output may be the power allocation ratio or operation command of each energy device. During this process, real-time constraints (such as device capacity and operation safety limits) need to be considered, and post-processing can be performed on the model output using policy clipping or normalization techniques. The entire decision-making process is generally completed in the controller, which requires high response speed to ensure quick adaptation to changes in load and energy supply.

[0060] The control instructions of this embodiment are shown as follows (in JSON format):

[0061] Fourth step: Strategy evaluation and explanation Input: The scheduling instructions output by the reinforcement learning model in the third step and their related context information. The context information includes the current system state (to understand what situation the strategy is making decisions for), historical scheduling data, system operation logs, and other multi-modal data (such as device conditions before this scheduling, recent weather reports, and operation and maintenance records). These input data types include model strategy output (numerical action vector) and text or image data for explanation and analysis (such as log text, thermal images, etc.).

[0062] Output: For each scheduled evaluation report and interpretation result, including: safety and rationality check results (such as whether to violate constraints, whether there are risk warnings), decision-making explainability analysis (such as policy explanation in natural language form), experience rule extraction results, and optimization suggestions. The specific form can be a structured text report or a rule list, such as: "In this strategy, the battery discharge is increased in advance, which can reduce the load burden of the gas turbine", or "According to the current ambient temperature, it is recommended to increase the use of cold storage tanks", etc. The output type is diverse, mainly in the form of readable text (policy explanation, warning explanation) and structured rules (key-value pairs or entry lists).

[0063] Core logic: Use large language models (LLM) to perform safety assessment and interpretation analysis on reinforcement learning strategies. LLM, based on the input strategy and system information, first performs safety review: judges whether the instructions comply with the operating constraints (such as device limitations, regulatory requirements), and marks potential risks. Then perform explainability analysis on the strategy: convert the complex decision-making logic inside the model into human-readable natural language descriptions, point out the key factors affecting the decision (similar to feature importance analysis), and extract implicit experience rules in system operation. For example, LLM can refer to historical scheduling logs and multi-modal data (operation and maintenance records, thermal imaging photos, etc.), extract the experience rule "prefer to use cold storage tank energy storage to smooth the cold load fluctuation in high temperature environment". Finally, LLM provides strategy fine-tuning suggestions based on identified model bias and business goals (such as suggesting to increase or decrease the output of a certain device, adjusting the time node, etc.), and generates contextual explanations and counterfactual analysis of the effects of the strategy (such as "if the light is brighter, the photovoltaic output can be increased, thereby reducing the number of gas turbine start-ups"). The core logic of this stage is to integrate the natural language reasoning capabilities of LLM and knowledge graph technology, enhancing the transparency and robustness of the scheduling strategy.

[0064] The LLM output of this embodiment is shown as follows (JSON format):

[0065] The above example shows the multi-dimensional output of LLM for a certain scheduling strategy, including: Safety check flag (safety_checks) is used to verify whether the strategy meets the system overload, voltage and thermal control constraints; executable constraint check (constraint_checks) is used to ensure that all continuous control quantities (such as photovoltaic output, battery power, unit and cold water unit power) are within the real-time available range or rated capacity; Natural language explanation (explanation) provides a textual description of the background of the scheduling decision; The extracted rules list (extracted_rules) enumerates typical control experiences distilled from historical data and multi-modal logs; the suggestions give feasible strategies for fine-tuning or system operation maintenance.

[0066] For example, the "explanation" field describes why discharging and adjusting the cold load are prioritized during peak load, and the "constraint_checks" explicitly states that all power commands in this strategy meet the available range of the device. The "extracted_rules" summarizes experiences such as "high light priority photovoltaic" and "peak smoothing discharge," and the "suggestions" proposes optimization measures such as "increase cold storage tank discharge" and "maintain SOC ≥ 20%."

[0067] Fifth step: Dynamic scheduling Input: Preliminary scheduling instructions of the reinforcement learning model and evaluation suggestions provided by the LLM. Specifically, it includes the original control action vector generated in the third step, and the safety suggestions, experience rules, and revision opinions output in the fourth step. These input data types include numerical commands and structured text (such as rule lists, numerical range warnings, etc.).

[0068] Output: The final executable scheduling commands issued to the energy router (or control execution unit). These commands are combined and adjusted to clearly indicate the actual operating parameters of each energy supply device and refrigeration device, such as the determined power generation, charging and discharging power, cold water unit start and stop state, and cold storage tank charging and discharging control, etc. The output form is a specific control signal or command format, usually a structured command required by the device control protocol, such as Modbus instructions, controller set value table, etc.

[0069] Core logic: The energy router control unit makes decision fusion and fine-tuning based on the input information: combines the scheduling strategy given by the reinforcement learning model with the suggestions of the LLM, resolves potential conflicts and optimizes the instructions. For example, if the LLM indicates that a certain output may cause device overload, the router will reduce the load of that device; if it proposes to increase the use of a certain energy, it will appropriately increase the corresponding output. In the fusion process, strategies such as weighted average, heuristic rules, or suboptimal decision trees can be used to combine the reinforcement learning output and experience rules to form the final command. The final generated command must meet the system safety, stability, and efficiency goals. Subsequently, the energy router will issue these commands to each physical device through the control bus or communication network to implement dynamic allocation of electrical load and cold load. This stage is the execution layer of the entire closed-loop control, requiring high real-time performance to achieve peak-valley filling scheduling and improve energy efficiency.

[0070] The scheduling command of this embodiment is shown as follows (JSON format):

[0071] In particular, it also includes a brief fine-tuning logic pseudo code: # Step 3: Get RL model output rl_cmd = get_rl_output(state) # returns dict with keys like "pv_power_set", "battery_power_set" etc # Step 4: Get LLM suggestions llm_suggestions = get_llm_suggestions() # returns object with lists of warnings and recommendations # Copy preliminary command final_cmd = rl_cmd.copy() # Adjust according to LLM rules: if "battery high temperature" warning, limit discharge power amplitude if "battery high temperature" in llm_suggestions.warnings: # Discharge power is negative, limit minimum discharge power to not exceed -30 kW final_cmd["battery_power_set"] = max(final_cmd["battery_power_set"], -30.0) # Increase photovoltaic output according to LLM recommendations if "increase photovoltaic output" in llm_suggestions.recommendations: # Increase by 5 kW, but not more than the current available photovoltaic power available_pv = get_current_pv_output() # get real-time available photovoltaic power final_cmd["pv_power_set"] = min(final_cmd["pv_power_set"] + 5.0, available_pv) # Send the final command, fine-tuned by LLM, to the energy router for execution send_to_router(final_cmd).

[0072] The above example demonstrates how the energy router utilizes both reinforcement learning strategies and the experience rules provided by the LLM to fine-tune and generate final execution commands. After five cycles of the five stages, the system continuously iterates and optimizes the scheduling strategy, ensuring safe, reliable, and energy-efficient operation while meeting load demands.

[0073] Therefore, the overall flow can be represented as the following pseudo code: def main_control_loop(): # Initialize modules scada = SCADAInterface() dt_env = DigitalTwinEnvironment(params) rl_agent = ReinforcementLearningAgent(model_path) llm = LargeLanguageModelInterface() router = EnergyRouterInterface() # Offline training phase (can be executed periodically) def offline_training(): # 1. Load historical real operation data real_data = scada.load_historical_data() # 2. Generate or load simulation data sim_data = dt_env.generate_simulation_data(real_data) # 3. Train reinforcement learning model rl_agent.train(real_data + sim_data) # 4. Save updated model rl_agent.save_model() # Online control main loop (real-time execution) while system_is_running(): #### First step: data collection #### # Read real-time raw data from SCADA raw_state = scada.read_realtime_data() # Verify, clean, and format state = preprocess_state(raw_state) #### Second step: (optional) offline training trigger #### if need_retraining(): offline_training() #### Third step: online decision making #### # Input current state into RL policy network to get preliminary action rl_cmd = rl_agent.infer(state) #### Fourth step: policy evaluation and explanation #### # Call LLM for safety and executability checks, generate explanation and suggestions llm_report = llm.evaluate_and_explain(state, rl_cmd) #### Fifth step: dynamic scheduling #### # Fine-tune RL output based on LLM suggestions final_cmd = apply_llm_suggestions(rl_cmd, llm_report) # Send final command to energy router for execution router.send_commands(final_cmd) # Wait for next scheduling interval wait_for_next_interval() # Data preprocessing function def preprocess_state(raw): # Noise removal, missing value imputation, unit conversion, time synchronization, etc. cleaned = clean_data(raw) return format_as_state_vector(cleaned) # Apply LLM suggested fine-tuning logic def apply_llm_suggestions(rl_cmd, report): cmd = rl_cmd.copy() # Safety constraint checks if not report.safety_checks['overload']: cmd = enforce_power_limits(cmd) # Device range checks cmd = enforce_device_capacity(cmd, report.constraint_checks) # Adjust according to empirical rules or suggestions for rule in report.extracted_rules: cmd = apply_extracted_rule(cmd, rule) for suggestion in report.suggestions: cmd = apply_suggestion(cmd, suggestion) return cmd # Whether retraining is needed (e.g. daily or when performance drops) def need_retraining(): # Can be based on model performance monitoring or fixed schedule return rl_agent.performance_below_threshold() or time_for_scheduled_retrain() # Start main loop if __name__ == "__main__": main_control_loop()

[0074] Through the above process loop execution, the system can constantly use real-time data and simulation data to optimize the scheduling strategy, to ensure that the energy supply system of the intelligent calculation center meets the load demand, at the same time, achieves the effect of safe and reliable operation and high efficiency and energy saving.

[0075] As shown in one embodiment of Figure 2 and Figure 3 , a control system module architecture diagram of the control method and system of the comprehensive energy supply device of the intelligent calculation center is Figure 2 , and Figure 3The application discloses a control method and system for a smart center comprehensive energy supply device, and relates to the technical field of energy supply. The control system of the smart center comprehensive energy supply device comprises an energy supply equipment, an SCADA monitoring unit, a reinforcement learning control unit, a large language model unit and an energy router control unit. The energy supply equipment is specifically divided into a power supply module and a cooling supply module. The power supply module comprises a gas turbine power generation device, an electrochemical energy storage device and an auxiliary power supply. The gas turbine power generation device generates electric energy by using a natural gas power device, and outputs steam to a steam-type refrigerator. The auxiliary power supply comprises a power grid and a photovoltaic power generation device. The electrochemical energy storage device is used for storing excess electric energy and supplementing power supply during a load peak. The cooling supply module comprises a steam-type refrigerator and an electrically-driven centrifugal refrigerator. The steam-type refrigerator is driven by high-temperature steam generated by the gas turbine power generation device or an internal combustion engine to perform a refrigeration cycle. The electrically-driven centrifugal refrigerator performs refrigeration by using system electric power. The cooling capacity output by the cooling supply module is delivered to the smart center through a closed circulation cooling water system (containing a closed cooling water device and a cold water storage tank). The control system architecture diagram is shown in the figure, and the connection relationship of the modules is shown. The SCADA monitoring unit is connected with the energy supply equipment, and is used for collecting equipment operation states and multi-modal data in real time, and writing the data into a real operation data buffer area. Meanwhile, a simulation data buffer area is used for storing historical energy supply and demand simulation data generated by a digital twin simulation environment. The reinforcement learning control unit is connected with the SCADA monitoring unit, and is used for obtaining data from the real operation data buffer area and the simulation data buffer area, performing mixed training on a reinforcement learning model, and generating preliminary scheduling instructions.

[0076] The large language model unit is connected with the reinforcement learning control unit and the SCADA monitoring unit, and is used for evaluating strategy safety, providing explainable analysis and optimization suggestions. The energy router control unit is connected with the large language model unit and the energy supply equipment, and is used for executing dynamic scheduling commands. The whole system cooperates through the units to realize intelligent control of the energy supply equipment.

[0077] In another embodiment, the control system comprises the following modules: energy supply equipment (such as a gas turbine power generation device, a steam-type refrigerator, an electrically-driven centrifugal refrigerator and the like), an SCADA data monitoring system, a real operation data buffer area, a simulation data buffer area, a digital twin simulation environment, a reinforcement learning control unit, a large language model decision evaluation and explanation unit and an energy router control unit. The SCADA data monitoring system is used for collecting operation states and multi-modal data (such as operation logs, temperature heat maps, structured reports and the like) of the energy supply equipment in real time, and writing the data into the real operation data buffer area. The simulation data buffer area is used for storing historical energy supply and demand simulation data generated by the digital twin simulation environment.

[0078] The reinforcement learning control unit obtains data from the real operation data cache and the simulation data cache, and performs hybrid training on the reinforcement learning model, thereby generating a preliminary energy scheduling strategy. The large language model decision evaluation and explanation unit performs safety verification and explainability analysis on the scheduling strategy output by the reinforcement learning unit, extracts experience rules, and proposes optimization suggestions. The energy router control unit receives scheduling instructions from the reinforcement learning and large language model, fine-tunes the final instructions, and issues them to each energy supply device for execution. The digital twin simulation environment generates simulation data based on the constructed system physical model or system model, providing samples for offline training; at the same time, the environment corrects the simulation model according to the actual operation results to improve the consistency of the simulation results with the actual situation.

[0079] The above modules form a closed-loop interaction through data caching and control signals: real-time operation data collected by SCADA continuously updates the real operation data cache, providing input for the reinforcement learning unit and the large language model; the scheduling strategy output by the reinforcement learning unit and the large language model is executed by the energy router control unit and the running results are fed back; the digital twin simulation environment corrects the model according to the new data and generates new training data. Through this closed-loop interaction, the modules work together to achieve adaptive optimization control of the energy supply of the intelligent computing center.

[0080] In another embodiment, the system includes power supply modules such as gas turbine power generation devices, external power grids, electrochemical energy storage devices, and photovoltaic power generation devices, as well as refrigeration modules such as steam refrigerators and electrically driven centrifugal refrigerators. The gas turbine power generation device uses a natural gas power device to generate electricity and simultaneously outputs steam to the steam refrigerator; the power grid and the photovoltaic power generation device serve as auxiliary power sources; the electrochemical energy storage device is used to store excess electricity and supplement power during peak load. In terms of refrigeration, the steam refrigerator is driven by high-temperature steam generated by the gas turbine power generation device or an internal combustion engine to drive the refrigeration cycle, and the electrically driven centrifugal refrigerator directly uses system power for refrigeration. The cold energy output by the above refrigeration equipment is delivered to the intelligent computing center through a closed circulation cooling water system (including a closed cooling water cooler and a cold water storage tank) to provide cooling for computing devices. Figure 3 The dashed arrows represent the directions of power supply energy flow and cooling energy flow in the system: the power supply energy flow points from the power generation device to the intelligent computing center, and the cooling energy flow points from the refrigeration equipment to the intelligent computing center.

[0081] The present application is not limited to the above-mentioned embodiments, and any changes in shape or material composition are allowed, as long as the structure design provided by the present application is adopted. Any variation of the present application shall be considered within the scope of protection.

Claims

1. A control method of a synthetic energy supply device of a cognitive center, characterized by, The application comprises the following steps: Real-time operation data of energy supply equipment and environmental sensors are collected by a SCADA system and preprocessed; Based on the collected data and simulation data generated by a simulation environment, an offline hybrid training is performed on a reinforcement learning model to generate an energy scheduling strategy; A real-time state is input into the trained reinforcement learning model to generate a preliminary scheduling instruction, a large language model is used for safety verification and explainability analysis, and the strategy is optimized; The preliminary scheduling instruction and the suggestion of the large language model are fused to generate a final scheduling command by an energy router control unit, which is then sent to the energy supply equipment for execution.

2. The control method of a comprehensive energy supply device of an intelligent calculation center according to claim 1, characterized in that, The reinforcement learning model generates a preliminary scheduling instruction by inputting a real-time state into the reinforcement learning model, considering real-time constraints, and outputting corresponding actions after forward calculation, and finally outputting the energy scheduling decision instruction predicted by the reinforcement learning model at the current time; the real-time state includes data from each energy device and the intelligent computing center at the current time, specifically including the output condition and available capacity of the energy device, and the power load and cooling load demand of the intelligent computing center. 3.The control method of a comprehensive energy supply device of an intelligent computing center according to claim 1 or 2, characterized in that, The safety verification and explainability analysis using the large language model include judging whether the instruction conforms to the operation constraints according to the input strategy and system information, labeling potential risks, then converting the complex decision logic inside the model into natural language description, pointing out the key factors affecting the decision, and extracting the implicit experience rules in system operation; the strategy optimization includes providing strategy fine-tuning suggestions based on the identified model bias and business objectives, and generating context explanations and counterfactual analysis of strategy effects; the system information includes the current system state and historical scheduling data as well as system operation logs.

4. The control method of claim 3, wherein, The fusion of the preliminary scheduling instruction and the suggestion of the large language model includes combining the scheduling strategy given by the reinforcement learning model with the suggestion of the large language model by the energy router control unit, using weighted averaging and heuristic rules as well as suboptimal decision tree strategy, combining the reinforcement learning output and experience rules to form the final command; the sending to the energy supply equipment for execution includes sending the command to each device through the control bus or communication network by the energy router to implement dynamic allocation of power load and cooling load.

5. The control method of claim 4, wherein, The simulation data is generated by a simulation environment, simulating the operation state and energy output and consumption of several energy devices and cooling source devices under different loads and environmental conditions, the input data includes multi-dimensional state vectors and possible action schemes as well as corresponding feedback; the offline hybrid training includes offline learning of the reinforcement learning algorithm based on input samples, the agent selects scheduling actions in the virtual environment through repeated trial and error, obtains rewards from the environment, and optimizes the policy network accordingly, finally outputs the trained reinforcement learning scheduling model, updated model parameters and performance indicators. 6.The control method of a comprehensive energy supply device of an intelligent computing center according to claim 4 or 5, characterized in that, The operation data includes the operation state of multiple energy devices and loads in the intelligent computing center, specifically including device state information, operation logs, real-time sensor data and visual perception data; the preprocessing includes time synchronization of data from different sources, analysis of transmitted data, and pushing them to real-time data cache or database.

7. A control system of a smart center integrated energy supply device, adopting the control method of the smart center integrated energy supply device according to any one of claims 1-6, characterized in that, The application comprises the following steps: The energy supply device comprises a power supply module and a cooling supply module; The SCADA monitoring unit collects real-time equipment operation status and multi-modal data, and is connected with the energy supply device; The reinforcement learning control unit trains a model based on the buffer data and outputs a scheduling strategy; The large language model unit evaluates the safety of the strategy, provides explainable analysis and optimization suggestions, and is connected with the reinforcement learning control unit and the SCADA monitoring unit; The energy router control unit executes dynamic scheduling commands and is connected with the large language model unit and the energy supply device.

8. The control system of a synthetic energy supply device of a wisdom center according to claim 7, characterized in that, The power supply module comprises a gas turbine power generation device and an electrochemical energy storage device, and an auxiliary power supply. The gas turbine power generation device generates electric energy by using a natural gas power device, and outputs steam to a steam-type refrigerator. The auxiliary power supply comprises a power grid and a photovoltaic power generation device. The electrochemical energy storage device stores excess electric energy and supplements power supply during peak load.

9. The control system of a synthetic energy supply device of a wisdom computing center according to claim 8, characterized in that, The SCADA data monitoring system collects real-time operation status and multi-modal data of each energy supply device, and writes them into a real operation data buffer. A simulation data buffer stores historical energy supply and demand simulation data generated by a digital twin simulation environment. The reinforcement learning control unit obtains data from the real operation data buffer and the simulation data buffer, and performs hybrid training on a reinforcement learning model, thereby generating preliminary scheduling instructions.

10. The control system of a synthetic energy supply device of a wisdom center according to claim 8, characterized in that, The cooling supply module comprises the steam-type refrigerator and an electrically driven centrifugal refrigerator. The steam-type refrigerator is driven by high-temperature steam generated by the gas turbine power generation device or an internal combustion engine. The electrically driven centrifugal refrigerator is driven by system power. The cooling capacity output by the cooling supply module is delivered to the intelligent computing center through a closed circulating cooling water system comprising a closed cooling water cooler and a cold water storage tank.

Images