Power supply and distribution control system based on virtual power plant

Abstract

The invention discloses a power supply and distribution control system based on a virtual power plant, and relates to the technical field of power distribution systems. The system comprises five core modules: a digital twin simulation control closed loop module constructs a complete closed loop of data acquisition, simulation calculation, control execution and model iteration, and realizes real-time synchronization of a physical system and a digital mirror image and accurate issuing of a control instruction; the multi-agent source-load-storage optimization module improves the source-load-storage collaborative scheduling efficiency through distributed agent collection and reinforcement learning; the multi-scene energy storage adaptive control module generates a scene-based energy storage strategy based on meteorological, load and electricity price multi-dimensional perception and model prediction control; and the multi-type energy storage ratio linkage module realizes multi-type energy storage cooperative operation based on an energy storage characteristic database and scenarized weight optimization. All the modules work cooperatively, the system response speed, the control precision and the power supply stability are effectively improved, and the system is suitable for industrial parks and other complex power supply and distribution scenes.

Description

Technical Field

[0001] This invention relates to the field of power distribution system technology, specifically to a power supply and distribution control system based on a virtual power plant. Background Technology

[0002] With the high penetration rate of distributed energy sources such as photovoltaics and wind power, virtual power plants have become a core technology for integrating decentralized energy resources and balancing electricity supply and demand. Traditional virtual power plant power distribution systems mostly rely on centralized dispatching models. Faced with the volatility of distributed energy sources and the diversity of user loads, their ability to perceive, dynamically predict, and quickly adjust the system's operating status in real time is gradually becoming insufficient. There is an urgent need to achieve accurate mapping of the system status through digital twins and improve dispatching flexibility through the synergy of multiple types of energy storage to meet the demand for highly reliable and efficient power distribution.

[0003] In view of the above, this application is hereby submitted. Summary of the Invention

[0004] The purpose of this invention is to provide a power supply and distribution control system based on a virtual power plant to solve the problems mentioned in the background art.

[0005] To solve the above-mentioned technical problems, the present invention provides a power supply and distribution control system based on a virtual power plant, including: a digital twin simulation control closed-loop module, a multi-smart entity source-load-storage optimization module, a multi-scenario energy storage adaptive control module, a multi-type energy storage ratio linkage module, and a virtual power plant-microgrid bidirectional switching module. The digital twin simulation control closed-loop module collects full-dimensional operation data of the virtual power plant through edge nodes, and constructs a digital mirror that is synchronized with the physical system at the millisecond level. The digital mirror maps the output of distributed power sources, load fluctuations, energy storage SOC values ​​and microgrid interface parameters in real time. At the same time, it directly converts the simulation results into real-time control parameters and sends them to other modules. It also receives the operation data after execution by other modules and updates the parameters of the digital mirror model in reverse, forming a complete closed loop of "data acquisition-simulation calculation-control execution-model iteration". The multi-type energy storage ratio linkage module includes an energy storage characteristic database and a dynamic ratio calculation unit. The energy storage characteristic database stores the response speed, capacity limit, and charge / discharge efficiency parameters of electrochemical energy storage, flywheel energy storage, and pumped hydro storage. The dynamic ratio calculation unit receives the scenario type and charge / discharge strategy output by the multi-scenario energy storage adaptive control module, calls the parameters in the energy storage characteristic database to calculate the coordinated ratio of different types of energy storage, and generates multi-type energy storage linkage control commands. The multi-intelligent energy source-load-storage optimization module, the multi-scenario energy storage adaptive control module, and the virtual power plant-microgrid bidirectional switching module all implement functions based on data provided by the digital mirror and feed back execution data to the digital mirror. The clear module functions and data interaction logic highlight the core role of the digital twin's "simulation-control" closed loop, and clearly present the technical details of the dynamic allocation of multiple types of energy storage. This solves the dual problems of the disconnect between digital twin and control and the insufficient scheduling capability of single energy storage in the existing technology. At the same time, it improves the system's scheduling response speed, control accuracy, and power supply stability, ensuring that the virtual power plant's power supply and distribution system is controllable and optimizable in all states.

[0006] Furthermore, the multi-agent source-load-storage optimization module includes distributed agent units and a reinforcement learning engine. The distributed agent units correspond to distributed power sources, energy storage devices, and user loads, respectively. Each agent unit collects real-time operating data of its corresponding object and uploads it to the reinforcement learning engine. The reinforcement learning engine uses real operating data output from digital mirrors and extreme scenario simulation prediction data as training samples to generate source-load-storage collaborative optimization instructions through iterative training. These instructions are used to adjust the output of distributed power sources, control the charging and discharging of energy storage, and guide user load response. With the support of multi-agent distributed decision-making and dual-dimensional training data, the module solves the problems of existing centralized scheduling being unable to adapt to diversified resource collaboration and low user response enthusiasm, thereby improving the efficiency of source-load-storage collaboration and increasing the total adjustable resources of the virtual power plant and the clean energy consumption rate.

[0007] Furthermore, the multi-scenario energy storage adaptive control module includes a multi-dimensional data acquisition unit and a scenario identification and decision-making unit. The multi-dimensional data acquisition unit collects meteorological early warning data, real-time user load data, and electricity market price signals. The scenario identification and decision-making unit extracts and analyzes features from the collected data, identifies the current scenario type of the power supply and distribution system, generates a matching adaptive charging and discharging strategy for energy storage based on the scenario type, and transmits the strategy to the dynamic ratio calculation unit of the multi-type energy storage ratio linkage module. Through multi-dimensional data perception and scenario-based strategy generation, the module solves the problem that existing energy storage strategies rely solely on day-ahead forecasts and cannot respond to emergencies in real time. This enables precise matching of energy storage charging and discharging behavior with actual system needs, provides a basis for dynamic ratio of multiple types of energy storage, and further improves energy storage utilization and equipment cycle life.

[0008] Furthermore, the virtual power plant-microgrid bidirectional switching module includes a bidirectional communication link and a switching control unit. The bidirectional communication link establishes a real-time data transmission channel between the virtual power plant and the microgrid, transmitting operational status data and resource demand signals from both sides. The switching control unit receives data on the impact range of grid faults and the power supply gap of the microgrid from the digital mirror output, controls the microgrid to switch between grid-connected and islanded states, and simultaneously schedules redundant resources of the virtual power plant to support the microgrid or calls upon redundant resources of the microgrid to supplement the virtual power plant. Through bidirectional communication and seamless switching control, the problems of unidirectional management and control of existing virtual power plants and microgrids and power interruption during faults are solved, the microgrid switching time is shortened, the power supply reliability is improved, and the total amount of adjustable resources of the virtual power plant is expanded.

[0009] Furthermore, the closed-loop implementation steps of the digital twin simulation control closed-loop module include: S1, edge nodes collect distributed power output, load fluctuations, energy storage SOC values, and microgrid interface parameters of the virtual power plant; S2, the collected data is transmitted to the digital mirror, and the digital mirror updates the physical system mapping state; S3, the digital mirror performs simulation calculations based on the updated state to generate control parameters; S4, the control parameters are distributed to other modules for execution; S5, other modules feed back the executed operating data to the digital mirror; S6, the digital mirror adjusts the model parameters according to the feedback data and returns to S1 to complete the iteration. Through clear closed-loop steps, the synchronization between the digital mirror and the physical system and the timeliness of control commands are ensured, solving the problem of disconnect between existing digital twin simulation and control steps, and further improving the system's operational stability and optimization efficiency.

[0010] Furthermore, the training steps of the reinforcement learning engine in the multi-agent source-load-storage optimization module include: S11, receiving real operating data and extreme scenario simulation prediction data output by digital mirror; S12, preprocessing the data, removing outliers and dividing it into training and validation sets; S13, training the reinforcement learning model based on the training set and verifying the model accuracy through the validation set; S14, if the model accuracy does not meet the preset requirements, adjusting the model parameters and returning to S13; S15, if the model accuracy meets the requirements, generating source-load-storage collaborative optimization instructions; through standardized training steps, the problems of single training data and insufficient accuracy of existing reinforcement learning models are solved, the rationality of collaborative optimization instructions is improved, and the efficient collaboration of source-load-storage resources is ensured.

[0011] Furthermore, the scenario identification steps of the multi-scenario energy storage adaptive control module include: S31, the multi-dimensional data acquisition unit collects meteorological early warning data, real-time user load data, and electricity market price signals; S32, the scenario identification decision unit extracts features from the collected data and filters out characteristic parameters such as load fluctuation amplitude, meteorological risk level, and electricity price range; S33, the characteristic parameters are matched with a preset scenario library to determine the current scenario type of the power supply and distribution system; S34, a corresponding energy storage adaptive charging and discharging strategy is generated according to the scenario type. Through standardized scenario identification steps, the accuracy of scenario type judgment is ensured, providing precise input for the dynamic allocation of multiple types of energy storage, and further improving the matching degree between energy storage strategy and system requirements.

[0012] Furthermore, the switching steps of the switching control unit in the virtual power plant-microgrid bidirectional switching module include: S21, receiving the grid fault impact range and microgrid power supply gap data output by the digital mirror; S22, determining the current operating status of the microgrid; S23, if the microgrid is in grid-connected state and the main grid is faulty, disconnecting the microgrid from the main grid, switching to an isolated grid state, and scheduling redundant resources of the virtual power plant to access the microgrid; S24, if the microgrid is in an isolated grid state and the power supply is insufficient, establishing a connection between the microgrid and the virtual power plant, switching to grid-connected state, and calling redundant resources of the microgrid to supplement the virtual power plant; S25, after the switching is completed, feeding back the switching result to the digital mirror and the bidirectional communication link; through clear switching steps, ensuring a smooth transition of the microgrid under different states, solving the problems of large voltage fluctuations and disordered operation during the existing switching process, and further improving power supply continuity and user satisfaction with electricity use.

[0013] Compared with the prior art, the beneficial effects of the present invention are: 1. By constructing a complete closed loop of "data acquisition - simulation calculation - control execution - model iteration" for digital twins, the limitation of traditional digital twins being used only for offline simulation is overcome: distributed acquisition at edge nodes enables data processing nearby, and incremental update technology ensures real-time synchronization between the physical system and the digital mirror. Then, through multi-timescale simulation and hierarchical command issuance, simulation results are directly converted into real-time control commands, effectively solving the core problem of the disconnect between digital twins and power supply and distribution control. At the same time, relying on a multi-agent distributed decision-making architecture and a dual-dimensional training mechanism of reinforcement learning, the traditional centralized scheduling is replaced, allowing photovoltaic, energy storage, and load agents to autonomously collect data and make decisions. Combined with the advantages of the DDPG algorithm in handling continuous actions, the system's response capability to energy fluctuations and the accuracy of source-load-storage coordination are significantly improved, ensuring the efficient and stable operation of the virtual power plant.

[0014] 2. Leveraging multi-scenario perception and multi-type energy storage collaborative innovation, it overcomes the limitations of traditional energy storage control: By collecting multi-dimensional data on meteorology, load, and electricity prices to extract core features, and combining typical scenario library matching and MPC rolling optimization, the energy storage strategy can dynamically adapt to different operating scenarios, avoiding the rigidity of fixed strategies; at the same time, it constructs an energy storage characteristic database to uniformly manage parameters of multiple types of energy storage, and adopts a scenario-based weighted optimization algorithm (allocating power according to efficiency, response speed, and capacity demand) to achieve complementary advantages between long-term electrochemical energy storage and rapid response flywheel energy storage, solving the functional shortcomings of single energy storage or experience-based matching. This not only improves energy storage utilization and extends equipment life, but also avoids power fluctuations caused by independent operation of multiple energy storage systems through PTP time synchronization and power monitoring, ensuring system stability.

[0015] 3. Overcoming the power supply challenges of traditional switching through a two-way microgrid switching strategy and achieving system-level coordination: Employing a "connect-then-disconnect" grid-connected to islanded grid transition strategy and a "smooth synchronization" islanded to grid-connected grid transition strategy, the system obtains real-time status information through a two-way communication link and makes decisions based on multi-source information fusion. In the event of a main grid failure, the system first connects to the virtual power plant's redundant resources before disconnecting the main grid; when the islanded grid's power supply is insufficient, it first synchronizes parameters before soft grid connection, avoiding switching interruptions and equipment impact. More importantly, each module achieves data interoperability and functional complementarity with a digital twin at its core. The digital twin provides global support, multi-agent decision-making, optimized configuration of energy storage modules, and power supply guarantee from switching modules, forming a synergistic effect that significantly improves the energy utilization efficiency, scheduling flexibility, and power supply reliability of the virtual power plant, adapting to the needs of complex scenarios such as industrial parks. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of a power supply and distribution control system based on a virtual power plant. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] Please see Figure 1 This invention provides a technical solution: a power supply and distribution control system based on a virtual power plant, which relies on the coordinated action of five core modules: a digital twin simulation control closed-loop module for control, a multi-intelligent energy source-load-storage optimization module for execution, a multi-scenario energy storage adaptive control module for adjustment, a multi-type energy storage ratio linkage module for resource scheduling, and a virtual power plant-microgrid bidirectional switching module for emergency response. These five modules work together to address complex power supply and distribution needs.

[0019] I. Application Scenario Setting: Taking a virtual power plant in an industrial park as the application object, the park is equipped with a 50MW rooftop photovoltaic power station distributed across three factory areas, a 30MW onshore wind farm on the open land north of the park, a 20MW / 100MWh lithium iron phosphate electrochemical energy storage system, and a 10MW / 1MWh flywheel energy storage system. A supporting 20MW microgrid covers eight manufacturing enterprises and two employee living areas. This system, through the coordinated operation of five core modules, solves problems such as complex scheduling, low energy storage utilization, and power outages during microgrid switching in traditional power supply and distribution systems, achieving efficient and stable power supply and distribution within the park.

[0020] II. Digital Twin Simulation Control Closed-Loop Module: By constructing a digital mirror that is synchronized with the physical system in real time, a complete closed loop of "data acquisition - simulation calculation - control execution - model iteration" is formed, breaking through the limitation of traditional digital twins being used only for offline simulation.

[0021] S1: Edge Node Full-Dimensional Data Acquisition: Traditional centralized data acquisition methods require transmitting distributed energy, energy storage, and load data to a remote control center, resulting in drawbacks such as high transmission latency, high data redundancy, and heavy server processing pressure. This step utilizes edge node-based data acquisition and preprocessing to shorten the data transmission path, reduce bandwidth consumption by redundant data, and provide support for real-time control. Specific technical methods are as follows: Edge computing gateways are deployed next to each equipment control cabinet, and the data acquisition frequency is divided according to the dynamic characteristics of the data: fast-changing parameters such as voltage and current are acquired at a frequency of 50Hz with a sampling interval of 0.02s; medium-changing parameters such as photovoltaic output, wind power output, and load power are acquired at a frequency of 1Hz; and slow-changing parameters such as energy storage SOC value and equipment temperature are acquired at a frequency of 0.1Hz. The acquired data is preprocessed by the edge gateways, and outliers such as sudden drops in photovoltaic output to 0 and sudden increases in energy storage SOC are removed using the Isolation Forest algorithm. After compression, the data is transmitted to the digital twin server through the park's fiber optic network.

[0022] Example: The edge gateway next to the photovoltaic inverter No. 1 in the park collects the measured output current value of 230A-232A every 0.02s and calculates the average output power of 48.5MW-49.2MW every 1s; the edge gateway of the energy storage station collects the SOC value of 51.8%-52.2% every 10s. If the energy storage SOC value suddenly jumps from 52% to 80% in one collection, the gateway determines it as a sensor malfunction, automatically removes the abnormal data, and only uploads the previously normal data. The data transmission delay is controlled within 100ms, significantly lower than traditional centralized data collection.

[0023] Compared to centralized data acquisition solutions based on publicly available documents, this system's edge-distributed acquisition reduces data transmission latency, edge preprocessing reduces invalid data transmission, and server data processing efficiency is improved, solving the problem that traditional solutions struggle to support millisecond-level real-time control.

[0024] S2: Digital Image Construction and Incremental Update: A single model architecture for the digital image can easily lead to ambiguous device state mapping and difficulty in fault location; a full update mode consumes a large amount of computing power and cannot meet real-time requirements. This step adopts a layered modeling and incremental update strategy to achieve a balance between accurate device state mapping and efficient updates. Specific technical methods are as follows: The digital mirror adopts a three-layer architecture: the device layer is based on mechanistic modeling, with photovoltaics using a modified single-diode model and energy storage using an equivalent circuit model; the system layer constructs an energy flow network based on graph theory, with devices as nodes and lines as edges; the environmental layer is based on data-driven correlation with external factors such as weather and electricity prices, with an update cycle of 1 hour. Updates employ an incremental update mechanism, comparing the collected data with the current mirror parameters every 100ms. Only parameters with changes exceeding thresholds (photovoltaic output change ≥1%) and energy storage SOC change ≥0.5%) are updated; parameters below the thresholds remain unchanged.

[0025] Example: In the digital mirroring equipment layer of the park, the photovoltaic power station model No. 1 is associated with an illumination of 800W / m. 2 The temperature was 26℃, and the simulated output was 48.8MW. In the system layer, the output of PV1 was transmitted to the energy storage station via a 35kV line, and then connected to the microgrid via a 10kV line. The environmental layer was associated with meteorological data showing "sunny, cloudless." After 100ms, the measured output of PV1 increased to 49.2MW, a change of 0.8%, which did not reach the 1% threshold, and the mirror image did not update this parameter; the solar irradiance increased to 820W / m². 2 The change was 2.5%, exceeding the threshold. The mirror only updated the illumination parameters and the corresponding photovoltaic simulation output of 49.5MW, and the update process took 50ms, which is much shorter than a full update.

[0026] Compared to publicly available documents, the static digital twin model updates every 10 minutes. This system's layered modeling improves the accuracy of device status positioning, and the incremental update mechanism reduces update latency, achieving millisecond-level synchronization between the physical system and the digital mirror.

[0027] S3: Multi-timescale simulation calculation: The control requirements of power supply and distribution systems vary across different time scales: real-time control requires millisecond-level response, intraday scheduling needs to balance accuracy and efficiency, and day-ahead planning requires high-precision optimization. Single-timescale simulation is insufficient to meet the needs of multiple scenarios. This step employs multi-timescale collaborative simulation to address different control problems specifically. Specific technical methods are as follows: A three-tiered simulation system is established: short-term (1 min-1 h), medium-term (1 h-1 d), and long-term (1 d-1 w) (one day to one week). The short-term simulation uses a simplified model to ignore equipment loss details, with a calculation interval of 100 ms, and is used for line overload early warning and load fluctuation mitigation. The medium-term simulation uses a medium-precision model to consider major equipment losses, with a calculation interval of 15 min, and is used for intraday energy storage charging and discharging plans and distributed energy output allocation. The long-term simulation uses a high-precision model to cover equipment aging and meteorological fluctuation details, with a calculation interval of 24 h, and is used for day-ahead dispatch planning. The core of the simulation is optimal power flow calculation, with objective functions including minimizing generation costs and minimizing load balance deviations. Constraints cover voltage range, line capacity, and equipment rated power.

[0028] Example: The park starts multi-timescale simulation at 9:00: Short-term simulation calculates the microgrid inlet line current every 100ms, predicting that the line current will reach 205A at 9:00:30, with a rated current of 200A, triggering an overload warning; Medium-term simulation calculates every 15 minutes, determining that from 10:00 to 11:00, the photovoltaic output will be 50MW, the load will be 18MW, and of the remaining 32MW, 20MW will be used for energy storage charging and 12MW will be used for grid connection; Long-term simulation was completed at 22:00 the previous day, and a basic plan for the day's "7:00-9:00 energy storage discharge and 12:00-14:00 energy storage charging" has been formulated.

[0029] Compared to the single-timescale simulation schemes in publicly available documents, this system improves short-term simulation response speed, medium-term scheduling economy, and long-term planning accuracy, achieving precise coverage of control requirements at different time scales.

[0030] S4: Hierarchical Control Parameter Issuance: Control commands are categorized by urgency into emergency commands (e.g., fault clearing, overload regulation) and routine commands (e.g., output fine-tuning, energy storage charging / discharging). Traditional single-channel transmission easily leads to queuing delays for emergency commands, missing optimal processing opportunities. This step employs a hierarchical channel and command verification mechanism to ensure the timeliness and effectiveness of command transmission. Specific technical measures are as follows: A dual-channel command transmission system is constructed: a high-speed dedicated channel uses direct fiber optic connection, with a transmission delay of ≤10ms for emergency commands, employing AES-128 encryption and CRC32 verification to ensure command integrity; the regular channel uses campus Ethernet, with a transmission delay of ≤100ms for regular commands, also configured with encryption and verification mechanisms. Feasibility checks are performed before commands are issued: emergency commands check the current status of the equipment, such as whether the energy storage SOC meets the discharge requirements; regular commands check system constraints, such as whether the total output exceeds the line capacity. If the check fails, recalculation is performed.

[0031] Example: At 9:00:00, a short-term simulation triggers a line overload warning. The digital mirror sends an emergency command via a high-speed dedicated channel: "Energy storage starts discharging 2MW at 9:00:00.1," reaching the energy storage inverter in 8ms. Simultaneously, a conventional command is sent via a regular channel: "PV output reduced by 5MW," reaching the PV inverter in 15ms. Energy storage and PV operate synchronously, and by 9:00:30, the line current stabilizes at 195A (rated 200A), preventing overload. If a traditional single-channel approach were used, the emergency command would have to wait for the conventional command queue, resulting in a delay exceeding 50ms, by which time the line would already be overloaded.

[0032] Compared to the single-channel transmission scheme for public documents, this system improves the response speed of emergency commands, the transmission efficiency of regular commands, and the effectiveness of commands.

[0033] S5: Differential Operation Data Feedback: In the traditional full data feedback mode, a large amount of redundant data, such as data with small deviations between actual and command values, consumes transmission bandwidth and server computing power, leading to increased feedback latency. This step uses a differential feedback mechanism, transmitting only data with deviations exceeding a threshold, reducing the data volume and improving feedback timeliness. Specific technical methods are as follows: After each module executes the instruction, the deviation between the actual value and the instruction value is calculated: if the absolute value of the deviation is ≤5%, only the deviation value and a timestamp precision of 1ms are transmitted; if the absolute value of the deviation is >5%, the full data, including device status, operating parameters, and anomaly description, is transmitted. A timestamp is appended to the feedback data to ensure synchronization with the digital mirror time and avoid timing discrepancies.

[0034] Example: An energy storage system executes a "discharge 2MW" command, but the actual discharge is 1.95MW with a deviation of -2.5% ≤ 5%, only reporting "deviation -0.05MW, time 9:00:00.1". Subsequently, a "charge 3MW" command is executed, but the actual charging is 2.8MW with a deviation of -6.7% > 5%, and full data is reported: "Actual charging power 2.8MW, SOC increased from 52% to 54.5%, battery temperature 25℃, time 9:10:00.0". The digital mirror quickly identifies charging anomalies based on the feedback data, triggering a parameter optimization process.

[0035] Traditional full-feedback systems transmit a large amount of data per transmission, while this system's differential feedback reduces the amount of data transmitted per transmission, shortens server feedback data processing time, and reduces anomaly identification latency.

[0036] S6: Adaptive Optimization of Model Parameters: Equipment aging and environmental changes can lead to increased deviations in digital image model parameters. For example, the temperature coefficient of photovoltaic power varies with the service life. Traditional manual parameter tuning takes six months to a year, making it difficult to adapt to dynamic parameter changes. This step employs an adaptive optimization algorithm to automatically iteratively update model parameters, ensuring simulation accuracy. Specific technical methods are as follows: The parameter optimization process is initiated daily at 2:00 AM during the off-peak period of grid load: Based on the feedback data of the day, the simulation error, such as the deviation between the simulated and actual photovoltaic output, is calculated. The gradient descent method is used to adjust model parameters, such as the photovoltaic power temperature coefficient and the energy storage charging and discharging efficiency, iterating until the error is ≤1%. After optimization, the simulation results are verified by comparing the measured data from the past hour. If the verification is successful, the model parameters are updated; otherwise, the iteration is restarted.

[0037] Example: The initial power temperature coefficient of the No. 1 photovoltaic power station in the park was set to -0.4% / ℃. After six months of operation, the simulated output was 5% higher than the actual output. During daily optimization at midnight, the algorithm gradually adjusted the coefficient to -0.42% / ℃, and after 10 iterations, the error decreased to 0.8%. After cleaning the photovoltaic panels, the efficiency improved, and the error rose to 3%. The algorithm automatically adjusted the coefficient to -0.41% / ℃, and the error recovered to 0.9%.

[0038] Compared to the manual parameter tuning methods in publicly available documents, this system has a shorter parameter optimization cycle, lower simulation errors, requires no manual intervention, and reduces labor costs.

[0039] III. Multi-Agent Source-Load-Storage Optimization Module: The multi-agent source-load-storage optimization module achieves autonomous collaborative decision-making of distributed energy, energy storage, and load through the collaboration of distributed agents and reinforcement learning engines, breaking through the limitations of traditional centralized scheduling in dealing with multi-device collaboration and dynamic scene changes.

[0040] S11: Distributed Intelligent Agent Data Acquisition and Upload: Traditional centralized intelligent agent acquisition modes suffer from single-point-of-failure risks; a failure at one acquisition point can lead to the loss of data across the entire system, and the data transmission path is long and has high latency. This step adopts a distributed intelligent agent architecture, dividing the responsibilities of intelligent agents according to device type, thereby improving system fault tolerance and data acquisition efficiency. Specific technical means are as follows: Three types of distributed intelligent agents are configured: photovoltaic intelligent agents covering photovoltaic power stations in three factory areas, energy storage intelligent agents covering electrochemical and flywheel energy storage, and load intelligent agents covering eight factories and dormitory areas, all implemented based on edge computing terminals. Each intelligent agent collects operational data of its corresponding equipment. The photovoltaic intelligent agent collects sunlight, power output, and temperature; the energy storage intelligent agent collects SOC and charging / discharging power; and the load intelligent agent collects power consumption and adjustable range. Data is uploaded to the park's data bus using a "publish-subscribe" model, and the reinforcement learning engine only subscribes to core data such as total photovoltaic power output, total energy storage SOC, and total load power.

[0041] Example: The park's photovoltaic smart system collects data every 5 seconds. Photovoltaic unit No. 1 has an output of 48MW and a solar irradiance of 800W / m². 2 Temperature 26℃; Photovoltaic output of No. 2: 32MW; Irradiance: 780W / m² 2At a temperature of 25℃, data is published to the data bus; the reinforcement learning engine only subscribes to data with a total photovoltaic output of 80MW, a total energy storage SOC of 52%, and a total load power of 20MW to avoid redundant data transmission. When the photovoltaic No. 2 intelligent agent experiences a temporary failure, the engine automatically retrieves backup data from the digital mirror to ensure data continuity.

[0042] Traditional centralized intelligent agent solutions suffer from high data loss rates and single-point failures. This system's distributed architecture reduces data loss rates and data transmission volume, while improving system scalability. Adding new devices only requires adding the corresponding intelligent agents.

[0043] S12: Data preprocessing, i.e., cleaning and partitioning: Outliers in the collected data, such as jump data caused by sensor malfunctions, and redundant features, such as highly correlated sunlight and photovoltaic output, can increase training bias and slow down convergence speed in reinforcement learning models. This step improves data quality and model training efficiency through data cleaning and partitioning. Specific technical methods are as follows: Data preprocessing consists of two steps: the first step uses the Isolation Forest algorithm to clean outliers and map data points to a high-dimensional space, and outliers are identified as anomalous and removed; the second step uses Principal Component Analysis (PCA) to reduce the dimensionality while retaining 95% of the information, and then divides the training set and validation set in a 7:3 ratio, with 70% of the data used for model training and 30% used for accuracy validation.

[0044] Example: There are 3 outliers in the photovoltaic data of the park with a light intensity of 0W / m. 2 However, for outputs of 50MW, temperatures of -10℃, and outputs of 100MW, the Isolation Forest algorithm fully identified and removed all instances. The original 5-dimensional features—illuminance, temperature, wind speed, output, and SOC—were reduced to 2-dimensionality using PCA, retaining 96% of the information. Of the 997 valid data points, 700 were used as the training set to learn "illuminance 800W / m²". 2 The model's prediction accuracy was verified by using 297 patterns, such as "corresponding to an output of 48MW".

[0045] Traditional data preprocessing using only thresholding methods results in low anomaly detection accuracy. This system's isolated forest algorithm improves anomaly detection accuracy, while PCA dimensionality reduction shortens model training time and reduces validation set prediction error.

[0046] S13: Reinforcement Learning Model Training: Traditional rule-based scheduling schemes, such as fixed-period energy storage charging and discharging, are difficult to adapt to dynamic scenarios such as weather changes and electricity price fluctuations. This step uses a reinforcement learning model to learn the optimal strategy through trial and error, improving scheduling flexibility and economy. Specific technical methods are as follows: A reinforcement learning model is constructed using the Deep Deterministic Policy Gradient (DDPG) algorithm, comprising an actor network and a critic network. The actor network generates continuous control commands, such as energy storage charging and discharging power and load adjustment, based on system state, illumination, electricity price, and load. The critic network evaluates the value of these commands, such as cost savings. Training data includes two types: 100,000 historical real-world data points from the past three months, and 50,000 extreme scenario data points generated by digital mirroring, such as sudden drops in wind power output and sudden load increases during typhoons. An experience replay mechanism is used to store the training data, and the target network undergoes a stable training process, iterating until the validation set accuracy reaches ≥95%.

[0047] Example: During model training, the input is "light intensity 800W / m". 2 Given a scenario with an electricity price of 0.7 yuan / kWh and a load of 18MW, the network outputs the instruction: "50MW of full photovoltaic power generation, 2MW of energy storage charging, and no load adjustment." The network of critics assesses this instruction as saving 2000 yuan per day, giving it a value score of 92 out of 100. After 1 million training iterations, the model's decision accuracy in the "sunny day with moderate electricity prices" scenario increased from 60% to 95%, and its decision accuracy in the "typhoon day" extreme scenario increased from 40% to 85%.

[0048] Compared to the publicly available Q-learning algorithm, which only supports discrete decision-making, the DDPG algorithm in this system supports continuous control commands, thus improving optimization accuracy; dual-dimensional training data enhances model robustness and convergence speed.

[0049] S14: Model Parameter Tuning: Hyperparameters of reinforcement learning models, such as learning rate and batch size, directly affect training efficiency and accuracy. Traditional manual parameter tuning relies on experience, resulting in numerous trial-and-error attempts and low efficiency. This step uses a Bayesian optimization algorithm to automatically find the optimal hyperparameters, reducing tuning costs. Specific technical methods are as follows: To minimize the validation set error, a Gaussian process is used to construct a surrogate model: five sets of hyperparameters are randomly tested, such as learning rates of 0.001, 0.005, and 0.01, and batch sizes of 32 and 64, and the corresponding validation set error is calculated. The surrogate model predicts the optimal hyperparameter combination based on the test results, iterating gradually until the error is ≤5%. The hyperparameter adjustment range is: learning rate 0.0001-0.01, batch size 16-128.

[0050] Example: The initial hyperparameters of the model are "learning rate 0.005, batch size 32", and the validation set error is 8%.

[0051] The Bayesian optimization was iterated five times: the first time the learning rate was adjusted to 0.004 with an error of 7%; the second time the batch size was adjusted to 64 with an error of 6%; the third time the learning rate was adjusted to 0.003 with an error of 4.5%; the fourth time the batch size was fine-tuned to 48 with an error of 4.2%; and the fifth time the optimal combination was confirmed with an error of 4.2%. The entire process took 2 hours, which is more efficient than manual parameter tuning.

[0052] Manually tuned hyperparameter combinations have low coverage, while Bayesian optimization improves coverage. The model accuracy corresponding to the optimal hyperparameter combination is improved, avoiding the subjective bias of manual tuning.

[0053] S15: Optimize Instruction Generation and Feasibility Verification: Instructions generated by the reinforcement learning model may exceed device limitations or system constraints; directly issuing them could lead to device damage or system failure. This step ensures instruction validity and system security through feasibility verification. Specific technical methods are as follows: Command verification is divided into equipment constraint verification and system constraint verification. Equipment constraint verification includes requirements such as energy storage commands needing to meet rated power and SOC range, and load commands needing to meet adjustable range. System constraint verification includes requirements such as total output needing to meet line capacity and voltage range. If the verification passes, the command is issued; if it fails, the model is returned to be regenerated, and the process iterates until the command meets the requirements.

[0054] Example: The model generates a "discharge 5MW from energy storage" command. Equipment limitation verification reveals that the rated energy storage power is only 4MW, and discharging 5MW at the current SOC of 10% would take 1.25 hours, indicating insufficient capacity. The model is then returned to its previous state. The adjusted command is "discharge 3MW," and verification is performed again: rated power, SOC, and line capacity are all met. After successful verification, the command is issued. After execution, the peak-valley load difference in the industrial park decreases from 10MW to 5MW, resulting in savings in daily electricity purchase costs.

[0055] Traditional system commands have low efficiency, but the verification mechanism of this system improves command efficiency; at the same time, it avoids equipment overload operation, extends the cycle life of electrochemical energy storage, and reduces equipment maintenance costs.

[0056] IV. Multi-scenario Energy Storage Adaptive Control Module: The multi-scenario energy storage adaptive control module realizes dynamic optimization and configuration of energy storage resources through multi-dimensional data perception and scenario-based strategy generation, breaking away from the limitations of traditional fixed energy storage strategies.

[0057] S31: Multi-dimensional Scene Data Collection: Single-dimensional data, such as load alone, is insufficient to accurately determine the operating scene. For example, load growth may originate from peak production periods or low electricity prices. This step collects multi-dimensional data to lay the foundation for accurate scene identification. Specific technical methods are as follows: Data is collected from three scenarios: meteorological data (sunlight, wind speed, and precipitation probability) is obtained from the provincial meteorological platform, with an update cycle of 1 hour; load data (power consumption in factories and dormitories) is collected through smart meters, with an update cycle of 15 minutes; and electricity price data (real-time and day-ahead prices) is obtained from the regional power trading platform, with real-time prices updated every 15 minutes and day-ahead prices released daily at 18:00 for the following day. Timestamps are appended to the collected data to ensure time-series alignment of multi-dimensional data.

[0058] Example: Multi-dimensional data collected in the park at 9:00 AM: Meteorological data: "Sunny day, sunshine 800W / m²" 2 The data includes: wind speed 3 m / s, no precipitation; load data: factory load 18 MW, up 1 MW from 8:45; dormitory load 2 MW, stable; total load 20 MW; electricity price data: real-time electricity price 0.7 yuan / kWh, up 0.1 yuan / kWh from 8:45; next day's highest price 0.9 yuan / kWh 18:00-20:00. All data is marked with a "9:00:00" timestamp to ensure temporal consistency of scene feature extraction.

[0059] Traditional scenario data is collected only with a load update cycle of 1 hour, resulting in a high scenario misjudgment rate. This system collects data from multiple dimensions, which reduces the scenario misjudgment rate. Meteorological data is acquired 1 hour in advance, increasing the lead time for strategy adjustments and significantly enhancing the ability to respond to emergencies.

[0060] S32: Scene Feature Extraction: Raw scene data is high-dimensional and redundant, and directly using it for scene recognition can easily lead to high model complexity and slow recognition speed. This step extracts core features to simplify the recognition process and improve judgment accuracy. Specific technical methods are as follows: Three core features were extracted: load fluctuation amplitude calculation, which measures the ratio of the difference between the maximum and minimum load values ​​within one hour to the average value, reflecting load stability; and meteorological risk level, with sunlight <300W / m². 2 Wind speeds >10 m / s are considered high-risk, 300 W / m 2 Light intensity ≤ 800W / m 2 Alternatively, a wind speed of 5 m / s ≤ 10 m / s is considered medium risk, and solar radiation > 800 W / m² is also considered medium risk. 2 Furthermore, wind speeds <5m / s indicate low risk; the electricity price range is as follows: real-time electricity price >0.9 yuan / kWh is high electricity price, 0.5 yuan / kWh ≤ real-time electricity price ≤0.9 yuan / kWh is medium electricity price, and real-time electricity price <0.5 yuan / kWh is low electricity price.

[0061] Example: Feature extraction for the 9:00 AM scenario in the industrial park: Maximum load of 20MW, minimum of 18MW, and average of 19MW within 1 hour, with a load fluctuation range of 10.5%; meteorological data corresponds to low risk; electricity price corresponds to medium electricity price. The final extracted features are "load fluctuation range of 10.5%, low meteorological risk, and medium electricity price", providing core basis for subsequent scenario matching.

[0062] Traditional feature extraction only uses the average load feature, resulting in low coverage. This system uses multiple features to improve coverage. At the same time, it can distinguish between "stable load growth" and "drastic load fluctuation" scenarios, improving the granularity of scenario recognition.

[0063] S33: Typical Scene Matching: Given the vast number of dynamic scenes, direct modeling and recognition are insufficient for efficient identification. This step improves recognition speed and accuracy by comparing and matching real-time features with typical scenes using a pre-set typical scene library. Specific technical methods are as follows: A typical scenario library is constructed, containing five core scenarios: weekday load fluctuation <15%, low meteorological risk, medium electricity price; weekday peak load fluctuation >20%, medium meteorological risk, high electricity price; weekend load <15MW, low meteorological risk, low electricity price; extreme weather load fluctuation >15%, high meteorological risk, medium electricity price; and electricity price arbitrage period with low electricity price, low meteorological risk, and load fluctuation <10%. Cosine similarity is used to calculate the similarity between real-time features and typical scenarios. Scenarios with a similarity of ≥90% are considered matching results; those with a similarity <90% are identified as mixed scenarios, and a weighted fusion strategy is used to generate a matching scheme.

[0064] Example: Comparison of real-time features at 9:00 AM in the park with typical scenario database: 95% cosine similarity with weekday flat-hour scenarios, 40% similarity with weekday peak-hour scenarios, and <30% similarity with other scenarios, ultimately matching the "weekday flat-hour" scenario. A certain real-time feature was "load fluctuation 18%, medium weather risk, medium electricity price," with similarities to both weekday flat-hour and peak-hour scenarios both <90%, thus classified as a mixed scenario, employing a weighted strategy of "flat-hour as the primary factor, peak-hour as a secondary factor." Comparison with existing technologies: Compared to the high mismatch rate of publicly available document rule matching schemes, this system uses cosine similarity matching to reduce the mismatch rate; the hybrid scene processing mechanism solves the "either / or" recognition limitation of traditional schemes, and improves the completeness of scene coverage.

[0065] S34: Energy Storage Adaptive Strategy Generation: The core objectives of energy storage differ across scenarios, making it difficult for fixed strategies to simultaneously address multiple needs. This step employs Model Predictive Control (MPC) to dynamically generate the optimal strategy based on the scenario objectives. Specific technical methods are as follows: The weights of the MPC objective function are determined based on the matching scenario: minimum cost during weekday flat periods (70%), load smoothing (30%), peak load smoothing during weekdays (70%), minimum cost (30%), power supply guarantee during extreme weather (80%), and minimum cost (20%). The MPC prediction time domain is set to 1 hour, and the control time domain is set to 15 minutes. Rolling optimization is performed every 15 minutes, executing only the strategy for the first 15 minutes. Subsequent optimizations are based on new data. Constraints cover the energy storage SOC range and charging / discharging power limits.

[0066] Example: The park matches the "weekday peak hours" scenario at 9:00 AM. The MPC objective function is "70% minimum cost + 30% load reduction". Predicted data for the next hour: Illuminance from 800W / m² 2 Increased to 850W / m 2 The photovoltaic (PV) output increased by 5MW, the load rose from 20MW to 21MW, and the electricity price increased from 0.7 yuan / kWh to 0.75 yuan / kWh. The optimized generation strategy is as follows: from 9:00 to 9:15, 2MW of energy storage is charged to utilize the excess PV output; from 9:15 to 9:30, 1MW is charged to accommodate the load increase; from 9:30 to 10:00, there is no charging or discharging, and the PV output just covers the load. After implementation, the energy storage SOC increased from 52% to 55%, load fluctuations were controlled within 3%, and daily costs were saved.

[0067] Traditional fixed-strategy energy storage has low efficiency, while the adaptive strategy of this system improves efficiency; MPC rolling optimization can cope with sudden scenarios, reduce strategy adjustment delay, and avoid load power shortage.

[0068] V. Multi-type energy storage ratio linkage module: The multi-type energy storage ratio linkage module achieves optimized allocation of energy storage resources through the synergy of different energy storage characteristics, solving the functional shortcomings of traditional single energy storage or simple superposition schemes.

[0069] S41: Energy Storage Strategy Reception and Verification: Energy storage strategies generated by multi-scenario modules may exceed the total capacity or power limits of multiple energy storage types. Direct execution may lead to energy storage overload or strategy failure. This step verifies the strategy to ensure its compatibility with energy storage capabilities. Specific technical methods are as follows: The system receives strategies from multiple scenario modules, including total charging and discharging power and duration. It then performs two types of checks: power check calculates the sum of the rated power of various energy storage types to ensure the strategy power is less than or equal to the total power; capacity check calculates the required capacity of the strategy to ensure it is less than or equal to the available energy storage capacity. If the checks pass, the system proceeds to the matching phase; otherwise, it feeds back to the multiple scenario modules to adjust the strategy, iterating until the checks pass.

[0070] Example: The multi-scenario module issues a "3MW charging from 9:00 to 10:00" strategy. Power verification reveals that the total rated power of 2.5MW from electrochemical energy storage (2MW) and flywheel energy storage (0.5MW) is less than 3MW, so the verification fails. After feedback, the multi-scenario module adjusts the strategy to "2.5MW charging" and verifies again: power 2.5MW = total power, capacity 2.5MW × 1 hour = 2.5MWh ≤ empty energy storage capacity 5MWh, and the verification passes.

[0071] Traditional systems lack a strategy verification step, resulting in low strategy efficiency. This system's verification mechanism improves strategy efficiency and avoids overload operation of energy storage, extending the lifespan of the energy storage converter.

[0072] S42: Energy Storage Characteristic Data Retrieval: Different types of energy storage exhibit significant differences in characteristics. Traditionally, parameters are stored in a scattered manner, such as being hard-coded into the code, making updates and maintenance difficult and prone to causing ratio deviations. This step constructs a characteristic database to uniformly manage parameters, providing an accurate basis for ratio calculations. Specific technical methods are as follows: A database of energy storage characteristics is built and managed using SQL Server, storing core parameters for each type of energy storage: electrochemical energy storage with a rated power of 2MW, charge / discharge efficiency of 90%, and response time of 100ms; flywheel energy storage with a rated power of 0.5MW, charge / discharge efficiency of 85%, and response time of 10ms. Parameters are updated monthly through on-site testing, such as changes in electrochemical efficiency over time, and are updated immediately after equipment maintenance or replacement. During ratio calculations, the required parameters are retrieved via SQL queries, ensuring the real-time nature and accuracy of the parameters.

[0073] Example: At 9:00 AM, the park implements a 2.5MW charging ratio. Parameters are retrieved via SQL query: electrochemical energy storage rated power 2MW, efficiency 90%, response time 100ms; flywheel energy storage rated power 0.5MW, efficiency 85%, response time 10ms. Based on these parameters, the charging ratio is determined as follows: priority is given to utilizing electrochemical energy storage due to its high efficiency and large capacity, while flywheel energy storage provides fast response to supplement remaining power and cope with fluctuations.

[0074] Traditional methods with hard-coded parameters have long update cycles, while this system's database management shortens update times; unified parameter calling avoids inconsistencies among parameters from multiple modules, improving the accuracy of proportion calculations.

[0075] S43: Calculation of Co-proportioning Ratio: Traditional empirical ratios of 80% electrochemical and 20% flywheel cannot adapt to changing scenarios. This step dynamically calculates the optimal ratio based on the scenario objectives, achieving precise matching between energy storage characteristics and scenario requirements. Specific technical methods are as follows: The target weights for allocation are determined based on the matching scenarios: weekday average efficiency weight 70%, cost weight 30%, short-term impact scenario response speed weight 80%, efficiency weight 20%, extreme weather capacity weight 90%, response speed weight 10%. Based on these weights, a comprehensive score for each type of energy storage is calculated, such as electrochemical efficiency score = efficiency × efficiency weight, and cost score = (1 - unit cost / highest unit cost) × cost weight. Power is then allocated according to the score ratio.

[0076] Example: In a 9:00 AM "weekday average" scenario within the park, the allocation target is "efficiency 70% + cost 30%": Electrochemical energy storage efficiency 90% × 70% = 63 points, unit cost 0.1 yuan / kWh (maximum 0.15 yuan / kWh × 30%) = 10 points, total 73 points; Flywheel energy storage efficiency 85% × 70% = 59.5 points, unit cost 0.15 yuan / kWh × 30% = 0 points, total 59.5 points. Based on the score ratio, allocate 2.5MW of power: 2MW of electrochemical energy storage and 0.5MW of flywheel energy storage, matching the scenario target.

[0077] Compared to the empirical matching schemes in publicly available documents, the scenario-based matching of this system improves the overall efficiency of energy storage, reduces the response time in short-term impact scenarios, and increases the capacity utilization rate in extreme weather.

[0078] S44: Generation and Issuance of Interlocking Control Commands: Independent startup of multiple types of energy storage can easily lead to fluctuations in total power, affecting grid stability. This step uses interlocking control to achieve synchronized operation of multiple energy storage systems, ensuring stable total power. Specific technical methods are as follows: When generating commands, a synchronization timestamp with a precision of 1ms is added, such as "9:00:00.0 Start charging, electrochemical energy storage 2MW, flywheel energy storage 0.5MW". The Precision Time Protocol (PTP) is used to ensure that the synchronization error of each energy storage clock is ≤1ms. After startup, the total power is monitored in real time. If the deviation exceeds 1%, such as the flywheel energy storage actually charging 0.6MW, a fine-tuning command is immediately issued, such as reducing the flywheel energy storage by 0.1MW, to ensure that the total power is stable within ±1% of the command value.

[0079] Example: At 9:00:00:00 in the park, electrochemical energy storage and flywheel energy storage start charging synchronously: electrochemical energy storage takes 100ms to increase from 0 to 2MW, and flywheel energy storage takes 10ms to increase from 0 to 0.5MW. Through PTP synchronization, the total power takes 100ms to increase from 0 to 2.5MW, with a maximum value of 2.52MW and a deviation of 0.8% ≤ 1%. During operation, the flywheel energy storage output fluctuates to 0.6MW, and the system immediately issues a "reduce 0.1MW" command, restoring the total power to 2.5MW, with fluctuations controlled within 0.5%.

[0080] Traditional independent control schemes have large total power fluctuations, while this system's linkage control reduces fluctuations and line voltage fluctuations; the real-time fine-tuning mechanism reduces power regulation delays and improves response speed.

[0081] VI. Virtual Power Plant-Microgrid Bidirectional Switching Module: The virtual power plant-microgrid bidirectional switching module achieves seamless switching between grid-connected and isolated grids through the "connect first, disconnect later" and "smooth synchronization" strategies, solving the power outage problem caused by the traditional "disconnect first, connect later" scheme.

[0082] S21: Real-time reception of operational status data: Switching decisions need to be based on real-time system status. Traditional data update cycles are long, easily leading to decision lag. This step uses high-frequency data reception to improve the timeliness of decision-making. Specific technical means are as follows: The system receives four types of real-time data from the digital mirror: the status of the main grid (good / bad), real-time power output from the main grid dispatch center and microgrid loads, power output from smart meters and microgrid power sources (PV and wind power), and power output from edge nodes and virtual power plants with available redundant resources. The data update cycle is 100ms, and a CRC32 checksum is appended to ensure data integrity; if the checksum fails, the data is retransmitted.

[0083] Example: At 10:00, the park receives real-time data: main grid status "bad", microgrid load 20MW, self-generated photovoltaic output 15MW with a 5MW shortfall, virtual power plant redundancy resources 8MW, and line status "on". Data is updated every 100ms. From 10:00:00 to 10:00:03, the main grid status is received four times consecutively as "bad", confirming that the main grid fault is real.

[0084] Traditional data update cycles are 1 second, resulting in long decision lags. This system updates data in 100ms, reducing decision lag. CRC32 verification reduces data error rates and avoids erroneous switching caused by false data.

[0085] S22: Microgrid Operation Status Assessment: Before switching, it is necessary to clarify the current status of the microgrid, such as grid-connected / islanded, sufficient / insufficient power supply, and equipment normal / faulty. Traditional single-status assessment can easily lead to incorrect switching strategies. This step ensures the accuracy of the strategy through multi-dimensional assessment. Specific technical means are as follows: The system determines three states: Connection state (microgrid and main grid switch positions, either connected or isolated); Power supply state (power output - load, positive for sufficient, negative for insufficient); Equipment state (operating status of key equipment such as inverters and switches, normal / faulty). Critical states are verified using a "two-out-of-three" criterion. For example, if there is a main grid fault, at least two of the three data types—main grid signal, line voltage, and current—must be confirmed to avoid misjudgment based on a single data point.

[0086] Example: Park status assessment at 10:00: Connection status "Grid-connected" (main grid switch closed); Power supply status "Insufficient" (15MW - 20MW = -5MW); Equipment status "Normal" (inverters and switches are fault-free). A main grid fault is confirmed by selecting two out of three criteria: "Main grid signal failure + line voltage 0kV + current 0A." If this is determined, a grid-connected to islanded grid switchover is required, and the 5MW redundant resources of the virtual power plant are dispatched to fill the gap. If the equipment status is "Faulty" (e.g., inverter damage), the equipment is repaired first before the switchover is performed.

[0087] Traditional single-state judgment schemes have a low success rate for switching, while this system improves the success rate through multi-dimensional judgment; the "two out of three" criterion reduces the false judgment rate of main network faults.

[0088] S23: Grid-connected to islanded grid handover, i.e., connection before disconnection: The traditional "disconnect the main grid first, then connect the virtual power plant" scheme will cause power outages. This step adopts a "connect first, then disconnect" strategy to achieve uninterrupted power supply handover. Specific technical means are as follows: The switching process consists of four steps: Step 1: Establish communication and power connection between the microgrid and the virtual power plant, and dispatch redundant resources to supplement the power supply gap, taking 50ms; Step 2: Gradually reduce the main grid power supply from the gap value to 0, taking 100ms; Step 3: Disconnect the main grid after the main grid power reaches 0, taking 20ms; Step 4: Stabilize the isolated grid by adjusting the virtual power plant's power supply to maintain a voltage of 380V±1% and a frequency of 50Hz±0.1Hz, taking 30ms. The entire process takes 200ms without any power interruption.

[0089] Example: The industrial park performs a grid-to-island switchover at 10:00:10:00:1 A virtual power plant connection is established, and 5MW of power is transmitted, with the microgrid voltage stabilizing at 380V; 10:00:0.1-10:00:0.2 The main grid power supply is reduced from 5MW to 0; 10:00:0.2 The main grid switch is disconnected; 10:00:0.2-10:00:0.23 The virtual power plant power supply is adjusted to 5.1MW to offset load fluctuations. Power supply remains uninterrupted throughout the process, the factory production line operates normally, with voltage fluctuations of 0.8% and frequency fluctuations of 0.05Hz.

[0090] Compared to the "disconnect first, then connect" solution in publicly available documents, which results in a longer power outage time, this system's "connect first, then disconnect" strategy eliminates power outage time, reduces factory production losses, lowers voltage and frequency fluctuations, and improves equipment operational stability.

[0091] S24: Islanded to Grid-Connected Switching (Smooth Synchronization): When a microgrid is operating in islanded mode, voltage and frequency differ from those of a virtual power plant. Direct grid connection would generate a large current surge. This step employs a smooth synchronization strategy to reduce the surge and ensure equipment safety. Specific technical measures are as follows: The switching process consists of four steps: Step 1: Measure the voltage difference, frequency difference, and phase difference between the microgrid and the virtual power plant, which takes 50ms; Step 2: Adjust the inverter on the synchronous microgrid side to ensure that the voltage difference is ≤2%, the frequency difference is ≤0.1Hz, and the phase difference is ≤5°, which takes 100ms; Step 3: Softly connect to the grid and slowly close the grid connection switch, increasing the power from 0 to the deficit value, which takes 200ms; Step 4: Stabilize the grid connection and adjust the power distribution, with the virtual power plant and the microgrid power supply working together, which takes 50ms.

[0092] Example: At 12:00, the photovoltaic output in the industrial park dropped to 10MW when switching from islanded to grid-connected, resulting in a power supply gap of 10MW: 12:00:0.1 Measured the voltage difference: 378V vs 380V, frequency 49.9Hz vs 50Hz, phase difference 3°; 12:00:0.1-12:00:0.2 Adjusted the microgrid voltage to 380V and frequency to 50Hz; 12:00:0.2-12:00:0.4 Slowly closed the grid-connected switch, increasing the power from 0 to 10MW; 12:00:0.4-12:00:0.45 Adjusted the power distribution: the virtual power plant supplied 7MW, the microgrid photovoltaic supplied 10MW, and the load was 17MW. The grid-connected current surge was 1.2 times the rated value, far lower than the surge from direct grid connection.

[0093] Traditional direct grid-connected solutions suffer from significant current surges and short switch lifespans. This system's smooth synchronization reduces these surges and extends switch lifespan; improved synchronization accuracy also shortens the voltage and frequency stabilization time after grid connection.

[0094] S25: Switching Result Feedback and Status Update: After switching, the result must be fed back to the digital mirror; otherwise, the digital mirror will continue to simulate in its original state, leading to deviations in subsequent instructions. This step ensures system state consistency through a feedback mechanism. Specific technical methods are as follows: After the switchover is complete, four types of information are fed back: switchover status (success / failure), key parameters (power outage time, voltage fluctuation, frequency fluctuation), current status (islanded / grid-connected), and resource allocation (power output distribution between the virtual power plant and the microgrid). Upon receiving the digital image, the microgrid model parameters are updated, such as changing the grid-connected status to islanded, to ensure that subsequent simulations match the current state with the commands.

[0095] Example: After the park successfully transitions from grid connection to islanded grid at 10:00 AM, the feedback information is as follows: Switching status "Successful", Outage time "0ms", Voltage fluctuation "±0.8%", Frequency fluctuation "±0.05Hz", Current status "Isolated", Resource allocation "Virtual power plant 5MW, Microgrid PV 15MW, Load 20MW". The digital mirror updates the microgrid model to islanded state, and subsequent simulations are calculated based on islanded grid. The instruction is adjusted to "Energy storage maintains SOC, avoids deep discharge".

[0096] Traditional systems lack a feedback mechanism, and deviations in digital mirror state lead to significant command errors. This system's feedback mechanism reduces these errors; feedback of key parameters provides a basis for optimizing switching strategies, resulting in continuous improvement in switching performance.

[0097] VII. Complete Example of Module Collaboration (Typical Operating Day of Virtual Power Plant in Industrial Park) 7:00-8:00: System Startup and Initialization: The digital twin module starts at 7:00, with edge nodes collecting initial data: 10MW photovoltaic, 20MW wind power, and 15MW load. A digital mirror is constructed, and long-term simulation is completed to formulate the daily plan of "12:00-14:00 energy storage charging and 18:00-19:00 energy storage discharging." The multi-agent module starts at 7:10, with distributed agents collecting and preprocessing data, which is then transmitted to the reinforcement learning engine. The engine loads the optimized parameters from the previous day and completes initialization. The multi-scenario module collects data at 7:20: 500W / m² illumination. 2 With a load of 15MW and an electricity price of 0.6 yuan / kWh, features were extracted and matched to the "weekday morning" scenario, and a "7:00-9:00 energy storage charging 1MW" strategy was issued.

[0098] 8:00-10:00: Flat Operation and Load Growth: At 8:00, the multi-type energy storage module receives the strategy, calculates the ratio of 0.8MW electrochemical energy storage and 0.2MW flywheel energy storage after verification, and issues a linkage command to start charging the energy storage. At 8:30, the digital twin module performs short-term simulation to predict microgrid line overload at 9:30 and sends an early warning to the multi-agent module. At 8:40, the multi-agent module generates the command "PV power reduction of 3MW, energy storage prepares to discharge 2MW", and issues it after verification. At 10:00, the virtual power plant-microgrid module receives main grid fault data, determines that the microgrid is "grid connected, 5MW short, equipment normal", executes "connect first, then disconnect" switching, and schedules the virtual power plant's 5MW redundant resources, ensuring no power outage.

[0099] 10:00-12:00: Isolated network operation and strategy adjustment: Multi-scenario module collects data at 10:30 AM. Illumination: 700W / m² 2 With a load of 19MW and an electricity price of 0.7 yuan / kWh, matching a "weekday flat period" scenario, the strategy was adjusted to "2MW of energy storage charging from 10:00 to 12:00". At 10:40, the multi-type energy storage modules received the new strategy and adjusted the ratio of electrochemical energy storage to 1.6MW and flywheel energy storage to 0.4MW, increasing the energy storage charging power to 2MW and the SOC from 52% to 58%. At 11:00, the digital twin module conducted a mid-term simulation prediction that the photovoltaic output would reach 50MW at 12:00, and sent a "prepare for more charging" command to the multi-agent module.

[0100] 12:00-14:00: Peak PV Power Generation and Energy Storage Charging: At 12:00, the multi-agent module generates a command for "50MW PV full power generation and 2.5MW energy storage charging," initially setting an overpower of 2.8MW. After adjustment and verification, the command is issued for execution. At 13:00, the multi-scenario module collects the electricity price at 0.65 yuan / kWh, matches the "electricity price arbitrage period," and maintains the "2.5MW charging" strategy. At 13:30, the multi-type energy storage module's SOC rises to 70%, and charging continues at 2.5MW.

[0101] 14:00-18:00: PV Power Decline and Grid Connection Recovery: The digital twin module performs a short-term simulation prediction at 14:00. At 16:00, PV output drops to 30MW, triggering an early warning to the multi-scenario module. At 14:30, the multi-scenario module adjusts its strategy to "14:00-18:00: Energy storage neither charging nor discharging, maintaining SOC." The virtual power plant-microgrid module, upon main grid recovery at 16:00, determines the microgrid to be "isolated, 2MW short, equipment normal," and executes a "smooth synchronization" switchover, successfully connecting to the grid.

[0102] 18:00-19:00: Peak Load and Energy Storage Discharge: At 18:00, the multi-scenario module collects data showing a load of 22MW and an electricity price of 0.9 yuan / kWh, matching the "weekday peak" scenario, and issues a strategy of "18:00-19:00 energy storage discharge of 2.5MW". At 18:10, the multi-type energy storage module receives the strategy, calculates the ratio of 2MW of electrochemical energy storage and 0.5MW of flywheel energy storage, and initiates energy storage discharge, reducing the SOC from 72% to 67%. At 18:30, the multi-agent module generates a command of "30MW of wind power generation, coordinated with energy storage discharge", which reduces the peak-valley load difference and saves on daily electricity purchase costs.

[0103] In summary, this system, through the collaborative innovation of five major modules, breaks through the technical bottlenecks of traditional virtual power plant power distribution systems. Its core innovations are reflected in four aspects: 1. Digital Twin Closed-Loop Control Innovation: Compared to centralized acquisition and static mirroring as described in publicly available documents, this system achieves a leap from "offline monitoring" to "online control" through reduced latency in edge acquisition, incremental update, and closed-loop iteration, improving real-time control response speed to within 100ms. 2. Multi-Agent Distributed Decision-Making Innovation: Compared to Q-learning discrete decision-making, this system adopts DDPG continuous decision-making to improve accuracy, dual-dimensional training data robustness, and Bayesian hyperparameter optimization efficiency, solving the problem of centralized scheduling's inability to handle multi-device collaboration and improving distributed resource scheduling efficiency. 3. Scenario-Based Energy Storage Strategy Innovation: Compared to rule matching as described in existing publicly available documents, this system achieves a transformation from "fixed mode" to "dynamic adaptive" energy storage strategies through reduced misjudgment rate in multi-dimensional acquisition, reduced mismatch rate in cosine similarity matching, and reduced response latency in MPC rolling optimization, improving energy storage efficiency. 4. Seamless Switching Technology Innovation: Compared to existing publicly available technologies that require interruption before reconnection, this system adopts a "reconnect-then-disconnect" approach, eliminating interruption time, reducing smooth synchronization impact, and improving the success rate of multi-state judgment. This solves the problem of power supply interruption during microgrid switching, improves power supply continuity, and extends equipment lifespan. Each module uses digital mirroring as its core to achieve data interoperability and functional complementarity, forming a system-level closed loop of "acquisition, simulation, decision-making, execution, feedback, and optimization." This achieves multi-module synergy, providing reliable technical support for virtual power plant power supply and distribution in complex scenarios such as industrial parks.

Claims

1. A power supply and distribution control system based on a virtual power plant, characterized in that: include: Digital twin simulation control closed-loop module, multi-smart body source-load-storage optimization module, multi-scenario energy storage adaptive control module, multi-type energy storage ratio linkage module and virtual power plant-microgrid bidirectional switching module; The digital twin simulation control closed-loop module collects full-dimensional operation data of the virtual power plant through edge nodes, and constructs a digital mirror that is synchronized with the physical system at the millisecond level. The digital mirror maps the output of distributed power sources, load fluctuations, energy storage SOC values ​​and microgrid interface parameters in real time. At the same time, it directly converts the simulation results into real-time control parameters and sends them to other modules. It also receives the running data after execution by other modules and updates the parameters of the digital mirror model in reverse, forming a complete closed loop. The multi-type energy storage ratio linkage module includes an energy storage characteristic database and a dynamic ratio calculation unit. The energy storage characteristic database stores the response speed, capacity limit, and charge / discharge efficiency parameters of electrochemical energy storage, flywheel energy storage, and pumped hydro storage. The dynamic ratio calculation unit receives the scenario type and charge / discharge strategy output by the multi-scenario energy storage adaptive control module, calls the parameters in the energy storage characteristic database to calculate the coordinated ratio of different types of energy storage, and generates multi-type energy storage linkage control commands. The multi-intelligent energy source-load-storage optimization module, the multi-scenario energy storage adaptive control module, and the virtual power plant-microgrid bidirectional switching module all implement functions based on data provided by the digital mirror and feed back execution data to the digital mirror.

2. The power supply and distribution control system based on a virtual power plant as described in claim 1, characterized in that: The multi-agent source-load-storage optimization module includes distributed agent units and a reinforcement learning engine. The distributed agent units correspond to distributed power sources, energy storage devices and user loads respectively. Each agent unit collects real-time operating data of the corresponding object and uploads it to the reinforcement learning engine. The reinforcement learning engine uses real operating data from digital mirror output and extreme scenario simulation prediction data as training samples to generate source-load-storage collaborative optimization instructions through iterative training. These instructions are used to adjust the output of distributed power sources, control the charging and discharging of energy storage, and guide user load response.

3. The power supply and distribution control system based on a virtual power plant as described in claim 2, characterized in that: The multi-scenario energy storage adaptive control module includes a multi-dimensional data acquisition unit and a scenario recognition and decision-making unit; The multi-dimensional data acquisition unit collects meteorological early warning data, real-time user load data, and electricity market price signals; the scene recognition and decision-making unit extracts and analyzes the features of the collected data, identifies the current scene type of the power supply and distribution system, generates a matching energy storage adaptive charging and discharging strategy based on the scene type, and transmits the strategy to the dynamic ratio calculation unit of the multi-type energy storage ratio linkage module.

4. The power supply and distribution control system based on a virtual power plant as described in claim 1, characterized in that: The virtual power plant-microgrid bidirectional switching module includes a bidirectional communication link and a switching control unit. The bidirectional communication link establishes a real-time data transmission channel between the virtual power plant and the microgrid, transmitting the operating status data and resource demand signals of both parties. The switching control unit receives the grid fault impact range and microgrid power supply gap data output by the digital mirror, controls the microgrid to switch between grid-connected and islanded states, and simultaneously schedules the virtual power plant's redundant resources to support the microgrid or calls upon the microgrid's redundant resources to supplement the virtual power plant.

5. The power supply and distribution control system based on a virtual power plant as described in claim 1, characterized in that: The closed-loop implementation steps of the digital twin simulation control closed-loop module include: S1, edge nodes collect distributed power output, load fluctuation, energy storage SOC value, and microgrid interface parameters of the virtual power plant; S2, the collected data is transmitted to the digital mirror, and the digital mirror updates the physical system mapping state; S3, the digital mirror performs simulation calculations based on the updated state to generate control parameters; S4, the control parameters are distributed to other modules for execution; S5, other modules feed back the executed running data to the digital mirror; S6, the digital mirror adjusts the model parameters according to the feedback data and returns to S1 to complete the iteration.

6. The power supply and distribution control system based on a virtual power plant as described in claim 2, characterized in that: The training steps of the reinforcement learning engine in the multi-agent source-load-storage optimization module include: S11, receiving real running data and extreme scenario simulation prediction data output by digital mirror; S12, preprocessing the data, removing outliers and dividing it into training and validation sets; S13, training the reinforcement learning model based on the training set and verifying the model accuracy through the validation set; S14, if the model accuracy does not meet the preset requirements, adjusting the model parameters and returning to S13; S15, if the model accuracy meets the requirements, generating source-load-storage collaborative optimization instructions.

7. The power supply and distribution control system based on a virtual power plant as described in claim 3, characterized in that: The scenario identification steps of the multi-scenario energy storage adaptive control module include: S31, the multi-dimensional data acquisition unit collects meteorological early warning data, real-time user load data, and electricity market price signals; S32, the scenario identification decision unit extracts features from the collected data and filters out characteristic parameters such as load fluctuation amplitude, meteorological risk level, and electricity price range; S33, the characteristic parameters are matched with a preset scenario library to determine the current scenario type of the power supply and distribution system; S34, a corresponding energy storage adaptive charging and discharging strategy is generated according to the scenario type.

8. The power supply and distribution control system based on a virtual power plant as described in claim 4, characterized in that: The switching steps of the switching control unit in the virtual power plant-microgrid bidirectional switching module include: S21, receiving the grid fault impact range and microgrid power supply gap data output by the digital mirror; S22, determining the current operating status of the microgrid; S23, if the microgrid is in grid-connected state and the main grid is faulty, disconnecting the microgrid from the main grid, switching to isolated grid state, and scheduling redundant resources of the virtual power plant to access the microgrid; S24, if the microgrid is in isolated grid state and power supply is insufficient, establishing a connection between the microgrid and the virtual power plant, switching to grid-connected state, and calling redundant resources of the microgrid to supplement the virtual power plant; S25, after the switching is completed, feeding back the switching result to the digital mirror and bidirectional communication link.

Images