Virtual power plant peak regulation optimization scheduling method and system, electronic equipment and medium

Abstract

The invention discloses a virtual power plant peak regulation optimization scheduling method and system, electronic equipment and a medium, and the method comprises the steps: obtaining multi-source data of distributed energy, energy storage equipment and a load side, and carrying out the cleaning and normalization processing; utilizing an LSTM neural network and an XGBoost model to carry out short-term and medium-and-long-term load prediction respectively; based on the prediction result and the resource state, generating a scheduling optimization strategy through a deep Q network (DQN) algorithm; according to an optimization strategy, distributed energy power generation, energy storage charging and discharging and load side response are dynamically managed, and multi-resource cooperative control is achieved; and through a feedback optimization module, DQN model parameters and strategy updating are automatically adjusted according to monitoring data, and the scheduling effect is optimized. The system comprises a data acquisition module, a data preprocessing module, a load prediction module, a scheduling decision module, a resource control module, and a monitoring and feedback optimization module. The method has the beneficial effects that the load prediction precision is improved, the resource scheduling is optimized, the operation cost is reduced, and the method is suitable for scheduling management of a multi-source energy system.

Description

Technical Field

[0001] The present invention relates to the field of electric power, and in particular to a virtual power plant peak load optimization scheduling method, system, electronic equipment and medium. Background Art

[0002] With the transformation of the global energy structure and the rapid development of renewable energy, the proportion of distributed energy in the power system has continued to increase, and a virtual power plant (VPP) consisting of multiple energy forms (such as solar energy, wind energy, energy storage systems, etc.) has gradually formed. Virtual power plants integrate and coordinate geographically dispersed distributed energy systems through information and communication technologies to achieve centralized control and optimized scheduling, and provide peak load regulation, peak load reduction and backup services for the power grid. However, due to the volatility and uncertainty of distributed energy, the load scheduling difficulty of the power system has increased, and the traditional centralized scheduling method is difficult to meet the dynamic response requirements of a high proportion of distributed energy.

[0003] In the process of peak load regulation of power system, accurate load forecasting is the basis for achieving optimized dispatching. Load forecasting is divided into short-term forecasting and medium- and long-term forecasting. Short-term load forecasting is usually used for real-time dispatching and response of power system, while medium- and long-term load forecasting provides support for long-term dispatching plan and power market transactions. With the widespread application of deep learning and machine learning algorithms, load forecasting models based on time series and ensemble learning have been widely studied and have shown excellent application prospects in the power industry.

[0004] In addition, deep reinforcement learning methods are expected to become an effective scheduling optimization tool for the dynamic coordination of multiple resources in virtual power plants. Reinforcement learning algorithms such as Deep Q Network (DQN) can theoretically generate optimized resource scheduling strategies based on power load demand, resource status, and cost factors. Such algorithms can continuously optimize decisions and adapt to different load fluctuations and demand responses in the face of environmental changes by continuously and adaptively adjusting their model parameters.

[0005] At present, some virtual power plant dispatching systems have achieved dynamic resource dispatching by combining distributed energy with load-side management. However, most of these systems fail to fully utilize the latest load forecasting and reinforcement learning technologies, and it is difficult to efficiently handle real-time load fluctuations, cost constraints, and resource allocation optimization in complex energy systems. In addition, the feedback optimization mechanism in the existing technology is usually relatively simple, and it is difficult to make timely adaptive adjustments based on the actual monitored dispatching effect, thus limiting the operation effect of the system.

[0006] Therefore, designing a virtual power plant peak-shaving optimization scheduling method and system that integrates short-term load forecasting, medium- and long-term load forecasting, real-time scheduling decision-making, resource coordination control, and feedback optimization adjustment can not only improve the system's responsiveness to load fluctuations, but also realize refined resource management and dynamic peak-shaving, which has important application value and broad market prospects. Summary of the invention

[0007] In order to solve the above technical problems, the present invention provides a virtual power plant peak load optimization scheduling method, system, electronic equipment and medium.

[0008] The present invention is achieved through the following technical solutions:

[0009] A virtual power plant peak load optimization scheduling method of the present invention comprises the following steps:

[0010] S1. Data collection

[0011] Through sensors, IoT terminals and data interfaces, distributed power generation status, load demand, environmental weather data and market price information are collected in real time. The real-time and integrity of multi-source data are ensured to provide a reliable data foundation for subsequent steps.

[0012] S2. Data Preprocessing

[0013] The collected multi-source data is cleaned, outliers are removed and normalized to ensure data consistency and accuracy. The pre-processed data is then imported into the load forecasting and scheduling optimization module to ensure the input data quality of subsequent algorithms.

[0014] S3. Short-term load forecasting

[0015] The Long Short-Term Memory (LSTM) neural network model is used to predict the load change trend within the next 1 to 24 hours. The short-term prediction results generated by this model provide accurate load demand data for fast-response real-time scheduling.

[0016] S4. Medium and long term load forecast

[0017] The XGBoost ensemble learning model is used to predict the power load trend in the next few days to weeks. The medium- and long-term forecast results provide support for the overall scheduling planning of the virtual power plant and help optimize resource allocation and reserve strategies.

[0018] S5. Real-time scheduling decision

[0019] According to the short-term and medium-term forecast results, the deep Q network (DQN) algorithm based on reinforcement learning is used to generate resource allocation and scheduling optimization strategies. This strategy comprehensively considers factors such as scheduling costs, resource constraints, and load balance to dynamically adjust the generation, storage, and load control of distributed energy.

[0020] S6. Resource coordination control

[0021] Execute dispatch strategies and use control module instructions to dynamically manage distributed energy systems (such as wind power, photovoltaics), energy storage equipment, and load-side resources. Ensure efficient collaboration among resources and achieve peak load balancing goals through multi-resource regulation.

[0022] S7. Operation effect monitoring

[0023] Through the monitoring module, the scheduling execution effect is collected and analyzed in real time, and the operation results of the strategy (such as response speed, economy and system balance) are evaluated. It ensures that the effect of the scheduling plan in actual operation is consistent with expectations, and finds possible scheduling deviations or optimization space.

[0024] S8. Feedback optimization and adjustment

[0025] According to the monitoring feedback data, the parameters of the deep Q network model are automatically adjusted, and the scheduling strategy is optimized and updated to cope with real-time changes and future needs. The scheduling method is continuously optimized, the robustness and adaptability of the system are enhanced, and the overall peak-shaving effect is improved.

[0026] The present invention provides a method for optimizing peak-shaving scheduling of a virtual power plant, which involves the steps of data collection, data preprocessing, load forecasting, real-time scheduling decision-making, resource coordination and control, operation effect monitoring and feedback optimization and adjustment. Deep learning and reinforcement learning algorithms are used to achieve optimized scheduling of distributed energy and energy storage equipment in the virtual power plant, so as to improve the peak-shaving response capability and operation efficiency of the system.

[0027] Furthermore, the above data preprocessing includes the following steps:

[0028] 1. Data cleaning

[0029] Missing value processing: Detect missing values ​​in the data set and process the records containing missing values. Common methods include:

[0030] Interpolation: Use the average of adjacent data or the trend of historical data to fill in missing values.

[0031] Mean replacement: replace missing values ​​with the mean of the same data feature.

[0032] Delete records: When there are too many missing values ​​or the impact on the analysis is small, delete the record directly.

[0033] Format standardization: standardize the format and units of the collected data, such as standardizing temperature data to degrees Celsius and time format to standard time.

[0034] 2. Outlier Removal

[0035] Statistical detection: detect outliers by calculating the mean, standard deviation and other statistical values ​​of the data. Mark or process extreme values ​​that exceed 3 times the standard deviation.

[0036] Logical rule detection: Rules are set according to the actual operating conditions and physical limitations of the virtual power plant. For example, the range of voltage, current or power should be within a reasonable operating range, and values ​​outside the range can be considered abnormal values.

[0037] Machine learning methods (such as the isolation forest algorithm): For high-dimensional data sets, machine learning anomaly detection algorithms can be applied to identify potential abnormal data points by building models.

[0038] 3. Normalization

[0039] Minimum-maximum scaling: scaling the data to a specific interval (such as 0-1), which is suitable for variables with a practical value range, such as load demand and electricity price. The formula is:

[0040]

[0041] Z-Score standardization: standardizes the data to a mean of 0 and a standard deviation of 1, which is suitable for data without a fixed range. The formula is:

[0042]

[0043] Among them, μ is the mean and σ is the standard deviation;

[0044] Specific normalization processing: For seasonal or periodic data (such as weather changes), you can consider using periodic normalization to map the values ​​to the sine function range of [-1,1] to retain the periodic characteristics of the data.

[0045] Furthermore, the above short-term load forecasting includes the following steps:

[0046] 1. Data preparation

[0047] Input data selection: Select key characteristic variables that have an impact on short-term load demand forecasting, usually including historical load data, weather information (such as temperature and humidity), time characteristics (such as hours and days of the week), and market price data.

[0048] Dataset partitioning: Divide the preprocessed data into training set, validation set, and test set, usually in a ratio of 80%, 10%, and 10%, to ensure the generalization ability of the model.

[0049] Time window partitioning: Divide the data into fixed time windows (such as load data for the past 24 hours) to generate time series input samples to ensure that the model can capture time-dependent characteristics.

[0050] 2. LSTM model construction

[0051] Model architecture design: Design the number of layers and nodes of the LSTM model. Usually it includes one or more LSTM layers, followed by a fully connected layer to output the predicted value.

[0052] Hyperparameter setting: Determine the hyperparameters of LSTM, including the number of hidden layer units, time step, learning rate, batch size, etc. These parameters will affect the training effect and prediction accuracy of the model.

[0053] Loss function selection: Mean Squared Error (MSE) is used as the loss function to measure the gap between the predicted value and the actual value to optimize the model performance.

[0054] 3. Model training

[0055] Forward propagation: Input the time series data of the training set into the LSTM model and calculate the predicted output.

[0056] Backpropagation and optimization: Use gradient descent method (such as Adam optimizer) to calculate the gradient of loss function with respect to model parameters, and gradually adjust model parameters to minimize the loss function.

[0057] Validation and early stopping: After each round of training, use the validation set data to evaluate the model performance and monitor whether overfitting occurs. Use the early stopping strategy to terminate the training when the validation set loss no longer decreases to prevent overfitting.

[0058] 4. Model Testing

[0059] Apply the trained LSTM model to the test set to evaluate its prediction accuracy on unseen data. Indicators such as root mean square error (RMSE) and mean absolute error (MAE) can be used to measure the prediction effect.

[0060] Compare the actual load data with the forecast results, analyze the forecast error and performance of the model to determine the feasibility of the model in practical applications.

[0061] 5. Generate prediction results

[0062] Short-term forecasting: Based on the test set or real-time input data, the trained LSTM model is used to make short-term forecasts of load demand within the next 1 to 24 hours.

[0063] Data output: The prediction results are output in time series format for use by subsequent scheduling decision modules.

[0064] 6. Result evaluation and model update

[0065] Regular evaluation: Based on the actual operation results of the system, the prediction accuracy of the model is regularly evaluated, and the model is retrained with the latest data to maintain the accuracy of the model.

[0066] Model parameter adjustment: When the load demand pattern or external environment changes significantly, the adaptability of the forecast can be improved by fine-tuning the model parameters or retraining the model.

[0067] The above steps constitute a complete short-term load forecasting process. Through the LSTM neural network model, the load demand fluctuations of the virtual power plant can be accurately captured, providing reliable forecasting data for scheduling optimization.

[0068] Furthermore, the above medium- and long-term load forecasting includes the following steps:

[0069] 1. Data preparation

[0070] Feature selection: Based on the needs of medium- and long-term forecasts, relevant feature variables are selected, such as historical load data, weather information (temperature, humidity, wind speed, etc.), calendar features (such as dates, holidays), electricity price data, etc.

[0071] Feature Engineering: Perform further feature engineering on the data to generate features that help improve prediction accuracy. For example, add lagging load features (such as the average load of the previous week) and sliding window mean to enhance the model's perception of load trends.

[0072] Dataset partitioning: Divide the data into training set, validation set, and test set. Since it is a medium- to long-term prediction, the data set can be divided in chronological order to ensure the model's generalization ability to future data.

[0073] 2. XGBoost model construction

[0074] Model initialization: Set the basic parameters of the XGBoost model, including the number of trees (n_estimators), tree depth (max_depth), learning rate (learning_rate), L1 and L2 regularization parameters, etc., to control the complexity of the model and the training speed.

[0075] Loss function selection: Use mean squared error (MSE) as the loss function to measure the deviation between the model's predicted value and the actual value.

[0076] 3. Feature Importance Analysis

[0077] Before training the model, key features are initially screened through feature importance analysis. XGBoost can rank the importance of each feature and help remove irrelevant or redundant features, thereby simplifying the model and improving training efficiency and prediction accuracy.

[0078] 4. Model Training

[0079] Iterative training: The training set data is input into the XGBoost model, and a series of decision trees are constructed through the iterative method of gradient boosting. Each tree is used to correct the prediction error of the previous round and gradually improve the overall prediction accuracy.

[0080] Validation and parameter adjustment: During the training process, the model performance is monitored through the validation set, and the hyperparameters (such as tree depth, learning rate, minimum number of sample splits, etc.) are adjusted using cross-validation or grid search to optimize the model effect and avoid overfitting.

[0081] 5. Model Testing

[0082] The trained XGBoost model is applied to the test set data to evaluate its prediction effect on unseen data. The prediction accuracy of the model is measured using indicators such as root mean square error (RMSE) and mean absolute error (MAE).

[0083] Analyze the difference between actual load and predicted load, and further adjust model parameters to ensure the stability of the model in medium- and long-term forecasts.

[0084] 6. Generate prediction results

[0085] Medium- and long-term forecasting: Based on real-time input data from actual applications and the trained XGBoost model, power load is predicted for the next few days to weeks.

[0086] Data output: The prediction results are output as time series data and provided to the subsequent scheduling optimization module.

[0087] 7. Result evaluation and model update

[0088] Regular evaluation: During the model application process, its prediction effect is regularly evaluated, and the model performance is analyzed based on the deviation between the medium- and long-term prediction results and the actual load data.

[0089] Model updating: When load patterns or other external conditions change significantly, retrain the model based on the latest data, or fine-tune model parameters to improve its adaptability to the new data.

[0090] These steps constitute the medium- and long-term load forecasting process of the XGBoost model. Through feature engineering and parameter optimization, XGBoost can accurately capture future power load trends and provide support for resource optimization scheduling of virtual power plants.

[0091] Furthermore, the above real-time scheduling decision includes the following steps:

[0092] 1. Define the scheduling environment

[0093] State space: Determines how to express the state of the environment. The state may include information such as the current generation of distributed energy, the power of energy storage equipment, load demand, and external grid prices.

[0094] Action space: defines available scheduling actions, such as adjusting the power generation of distributed energy, controlling the charge and discharge rate of energy storage equipment, and adjusting the controllable load on the load side.

[0095] Reward function: Design a reward function to measure the pros and cons of different scheduling schemes. Usually, minimizing scheduling cost, balancing energy consumption, and satisfying resource constraints are used as reward indicators to guide the DQN algorithm to find the optimal strategy.

[0096] 2. DQN model construction

[0097] Network architecture design: Build a deep Q network, including an input layer (for receiving state information), a hidden layer (commonly used multi-layer fully connected layers or convolutional layers to extract features), and an output layer (corresponding to the Q values ​​of different scheduling actions).

[0098] Experience replay mechanism: Set up an experience replay pool to store information such as the state, action, reward, next state, etc. of each execution, so as to facilitate subsequent small-batch random sampling training, break the time correlation of data, and improve learning efficiency.

[0099] Target network mechanism: Establish a target Q network and regularly copy the parameters of the main network to the target network to stabilize the training process and avoid excessive Q value fluctuations.

[0100] 3. Model training

[0101] Initialization: Initialize the system state to the current power system state, including the state of power generation, load and energy storage equipment.

[0102] Q-value update: In each step, the current state is collected, and the optimal action (that is, the action that maximizes the Q-value) is selected, the immediate reward is calculated according to the reward function, and the parameters of the deep Q network are updated.

[0103] The Q value update formula is:

[0104]

[0105] Where α is the learning rate, r is the immediate reward, γ is the discount factor, and Q(s′,a′) is the maximum Q value of the next state;

[0106] Exploration and Utilization: Using the ε-greedy strategy, a certain proportion of actions are randomly selected (exploration), and the rest are actions corresponding to the current optimal Q value (utilization) to improve the algorithm's exploration ability and scheduling effect.

[0107] 4. Real-time decision execution

[0108] According to the trained DQN model, the Q value of each possible scheduling action is calculated in real time, the action with the maximum Q value is selected, and the resource allocation and scheduling optimization strategy is generated.

[0109] Convert strategies into control instructions to dynamically adjust distributed energy generation, energy storage charging and discharging rates, and load regulation to ensure optimal coordination of various resources under different conditions.

[0110] 5. Monitoring of strategy effectiveness

[0111] Continuously monitor the execution effect of the scheduling strategy, record the status, action, reward and final effect of each decision, and store them in the experience replay pool for subsequent optimization and strengthening of the model.

[0112] Analyze the actual effect and deviation of scheduling. If there is a large deviation, feedback will be given to the training module to help the DQN model to optimize itself.

[0113] 6. Strategy update and optimization

[0114] Dynamic adjustment: Based on load forecasts, environmental changes, and real-time feedback data, the parameters of the DQN model are continuously adjusted to optimize the accuracy and adaptability of the scheduling strategy.

[0115] Periodic retraining: As the operation and load patterns of the power system change, the DQN model is regularly retrained using the latest data to ensure that it always meets the scheduling requirements of the current system.

[0116] Through the above steps, the real-time scheduling decision-making method based on the DQN algorithm can generate an efficient resource allocation strategy, dynamically adapt to different power load demands, and provide support for the efficient operation of virtual power plants.

[0117] Furthermore, the resource coordination control includes the following steps:

[0118] 1. Receive scheduling optimization strategy

[0119] The control module receives the scheduling optimization strategy generated by DQN in the previous step, including the generation instructions of distributed energy, the charging and discharging plans of energy storage equipment, and the response requirements on the load side.

[0120] 2. Distributed energy control

[0121] Power generation regulation: Dynamically adjust the power generation of distributed energy (such as photovoltaics, wind power, etc.) according to the dispatch strategy. When weather conditions or resource fluctuations affect power generation, the control system automatically reduces or increases the power generation to maintain system balance.

[0122] Grid connection management: When the load demand changes significantly, the control module adjusts the grid connection status of the distributed energy to ensure stable power output; and quickly disconnects or isolates some distributed power sources in the event of overload or fault to avoid affecting system safety.

[0123] 3. Energy storage equipment control

[0124] Execution of charging and discharging strategies: According to the dispatching strategy, energy storage equipment (such as battery energy storage, pumped storage, etc.) is controlled to charge or discharge at different time periods. Charging is usually arranged when the load is low or when there is excess renewable energy, while discharging is arranged when the load is peak or when the electricity price is high to reduce load pressure.

[0125] Status monitoring and safety control: Real-time monitoring of the power, temperature and health status of energy storage equipment to ensure that it operates within a safe range. When the device is close to full or the power is low, the charge and discharge rate is automatically adjusted to avoid damage to the device due to overcharging or over-discharging.

[0126] 4. Load side control

[0127] Load peak shaving: During peak load periods, load peak shaving is achieved by controlling adjustable loads (such as industrial cooling systems, heating equipment, etc.) to delay or run in different time periods; for interruptible loads, they are temporarily shut down or the power is reduced according to policy requirements.

[0128] Demand response: For controllable load resources, such as smart appliances or adjustable electrical equipment, price signals or incentives are used to guide users to increase electricity consumption during off-peak hours or reduce electricity consumption during peak hours, thereby achieving the effect of demand response.

[0129] 5. Coordinate feedback mechanism

[0130] Feedback data collection: Through sensors and monitoring equipment, the operation data of each resource is collected in real time, including power generation, energy storage status, load changes and other information. The feedback data is transmitted to the scheduling module to ensure the execution effect of the real-time monitoring scheduling strategy.

[0131] Abnormality detection and processing: When abnormalities are detected during the control execution process (such as large load fluctuations or abnormal energy storage equipment), the dispatching system is promptly notified through the feedback mechanism to regenerate the dispatching strategy or adjust the current control plan to ensure system stability.

[0132] 6. Real-time optimization and adjustment

[0133] Dynamic adjustment strategy: In the actual execution process, the control strategy is dynamically fine-tuned according to feedback data and system operation status. For example, the discharge power is reduced when the energy storage device is insufficient, or the critical load is prioritized when the load demand surges.

[0134] Strategy effect evaluation: Regularly evaluate the execution effect of the control strategy, including the synergistic efficiency of power generation, energy storage and load and the overall load balance of the system. Optimize the control model based on the evaluation results to improve the applicability of subsequent scheduling strategies.

[0135] Resource coordination control forms a closed-loop feedback and dynamic optimization mechanism by controlling distributed energy, energy storage equipment and load-side resources, ensuring the effective implementation of scheduling strategies and achieving the goals of efficient peak regulation and resource coordination.

[0136] Furthermore, the above operation effect monitoring includes the following steps:

[0137] 1. Initialize the monitoring module

[0138] Parameter configuration: According to the scheduling optimization strategy and system requirements, configure the key monitoring indicators of the monitoring module, such as distributed energy generation power, energy storage equipment charging and discharging status, load response, etc.

[0139] Monitoring frequency setting: According to the requirements of system dynamic changes, set the collection frequency of various types of data. For example, power generation and load response can be collected every minute, while energy storage device status can be collected every 5 minutes to balance real-time performance and data processing load.

[0140] 2. Real-time data collection

[0141] Status data acquisition: Through sensors, smart meters and distributed controllers, the status data of distributed energy, energy storage equipment and load side, including voltage, current, power, charge and discharge rate, etc., are collected in real time.

[0142] Environmental data collection: For environmental factors that affect the execution of scheduling, such as weather information (temperature, humidity, solar radiation, etc.), market electricity prices, etc., external data is collected regularly to ensure that the scheduling strategy adapts to changes in external conditions.

[0143] 3. Data Preprocessing

[0144] Data cleaning: Remove abnormal values ​​or erroneous data from monitoring data to ensure data accuracy. For example, eliminate extreme values ​​or sudden changes caused by equipment sensor failure.

[0145] Data smoothing and filtering: Use methods such as moving average or low-pass filtering to smooth out the noise in the data and improve the accuracy of subsequent analysis.

[0146] Data formatting: Storing or transmitting data in a unified format to provide a standardized data basis for further analysis and evaluation.

[0147] 4. Scheduling effect analysis

[0148] Real-time and planned deviation analysis: Compare the monitoring data with the expected dispatch plan to analyze the deviation between the actual execution and the plan. For example, calculate the deviation between the actual load demand and the predicted load, and the difference between the actual charge and discharge of the energy storage equipment and the planned amount.

[0149] Key Performance Indicator (KPI) evaluation: Based on the key performance indicators defined by the system (such as dispatch response speed, load peak shaving effect, energy storage utilization rate, etc.), calculate the performance of the current dispatch strategy and evaluate whether its effect meets the target.

[0150] Abnormal detection and alarm: When significant deviations or abnormal conditions are found in the scheduling execution (such as load overload, abnormal charging and discharging of energy storage equipment, etc.), the alarm is immediately triggered and the abnormality is recorded for subsequent analysis and adjustment.

[0151] 5. Generation of effect evaluation report

[0152] Generate operation effect report: Generate a detailed effect evaluation report based on the data analysis results. The report includes the actual data of scheduling execution, deviation analysis, abnormal records, KPI results, etc., providing a basis for system decision-making and strategy optimization.

[0153] Trend analysis and forecasting: Based on historical operating data, analyze the changing trends of indicators such as load demand, power generation, and energy storage utilization to provide reference for future scheduling strategies.

[0154] 6. Feedback and strategy optimization

[0155] Feedback to the scheduling module: Feedback the results of the operation effect monitoring to the scheduling optimization module. When significant deviations or continuous anomalies are detected, it is recommended that the scheduling module adjust the strategy to improve the real-time response capability of the system.

[0156] Adaptive optimization: Automatically adjust the monitoring frequency, parameter range, etc. of the monitoring module according to feedback data to adapt to the real-time demand changes of the system and ensure that the monitoring effect is consistent with the scheduling strategy.

[0157] Through the above steps, the operation effect monitoring module can evaluate the execution effect of the scheduling strategy in real time and accurately, provide data support for strategy optimization and system stability, and effectively improve the dynamic adaptability of virtual power plant scheduling.

[0158] Furthermore, the above feedback optimization adjustment includes the following steps:

[0159] 1. Data feedback collection

[0160] Collect key feedback data: Collect data feedback results from the operation effect monitoring module, including deviation analysis of scheduling execution, abnormal conditions, load fluctuations, energy storage equipment utilization, etc.

[0161] Extract model optimization information: Extract indicators related to DQN model performance from feedback data, such as load forecast error, scheduling response speed, peak load regulation effect, etc., to determine whether the model needs to be optimized and adjusted.

[0162] 2. Model Performance Evaluation

[0163] Calculate strategy benefits and costs: Calculate the overall benefits (such as peak shaving effect) and costs (such as resource consumption) of the current scheduling strategy based on feedback data as evaluation indicators of DQN model performance.

[0164] Determine strategy adaptability: Compare the current scheduling effect with environmental requirements, evaluate the adaptability of the existing DQN model in dealing with load demand fluctuations and environmental changes, and determine whether optimization is needed.

[0165] 3. Parameter adjustment and adaptive update

[0166] Learning rate adjustment: According to the frequency of environmental changes and the amplitude of demand fluctuations in the feedback, the learning rate of the DQN model is adjusted so that the model can adapt to new data faster or more robustly.

[0167] Exploration rate adjustment: By increasing or decreasing the exploration rate ε of the ε-greedy strategy, the exploration and utilization of the model can be balanced to improve the adaptability of the model in a dynamic environment.

[0168] Q-value update parameter optimization: According to the deviation analysis and policy response speed, the discount factor γ of the DQN model is adjusted to ensure that the model strikes a balance between long-term benefits and short-term responses.

[0169] 4. Strategy Retraining

[0170] Experience replay pool update: The latest feedback data is stored in the DQN experience replay pool to ensure that the model learns the latest system dynamic changes.

[0171] Mini-batch training update: Randomly extract new and old data samples from the experience replay pool, perform mini-batch training, and update the DQN model parameters to adapt them to the current system environment.

[0172] Target network synchronization: After a training iteration, the parameters of the main network are copied to the target network to stabilize the model update process.

[0173] 5. Scheduling strategy update generation

[0174] New strategy generation: The updated DQN model is used to generate new scheduling optimization strategies, including resource allocation, energy storage charging and discharging plans, load-side response measures, etc., to better adapt to the current load demand and resource status.

[0175] Multi-scenario testing and evaluation: Test new strategies in a simulation environment, simulating various environments and demand fluctuations to ensure that the strategy is robust and stable.

[0176] 6. Strategy effect verification and iterative optimization

[0177] Monitoring of strategy execution effect: After applying a new scheduling optimization strategy, continuously monitor its execution effect and collect feedback data; focus on monitoring the effectiveness of the new strategy in load peak shaving, resource coordination, and abnormal response.

[0178] Iterative optimization: When the actual effect does not meet expectations, the next round of parameter adjustment and retraining of the DQN model is automatically triggered to achieve a feedback loop and continuously optimize the performance of the scheduling strategy.

[0179] Through the above steps, the feedback optimization and adjustment module can realize adaptive optimization based on monitoring feedback, ensuring that the DQN model and scheduling strategy always adapt to environmental changes and continuously improve the performance of the scheduling method of the virtual power plant.

[0180] A virtual power plant peak load optimization dispatching system, comprising:

[0181] 1. Data acquisition module, used to collect multi-source data of the virtual power plant; the above data acquisition module includes:

[0182] Distributed energy generation data acquisition unit, used to collect real-time operating data such as power generation power, voltage, current, etc. of each distributed energy source;

[0183] Energy storage device status data acquisition unit, used to monitor information such as the power, charge and discharge rate, temperature and health status of the energy storage device;

[0184] The load-side demand data acquisition unit is used to obtain the power demand, load fluctuation and demand response status of each load device;

[0185] Environmental data acquisition unit, used to obtain external data such as weather information (temperature, humidity, light intensity, etc.) and electricity market prices.

[0186] 2. Data preprocessing module, used to process the multi-source data collected by the data acquisition module, including:

[0187] Data cleaning unit, used to remove noise data and outliers to ensure data accuracy;

[0188] Outlier detection and removal unit, used to mark and remove extreme data that exceeds the preset range;

[0189] The data normalization unit is used to normalize the data to ensure the consistency of the data.

[0190] 3. Short-term load forecasting module, used to forecast short-term load demand in the next 1 to 24 hours based on the data processed by the data preprocessing module, including:

[0191] LSTM model building unit, used to build and train Long Short-Term Memory (LSTM) neural network models to adapt to the nonlinear and time series characteristics of historical load data;

[0192] The short-term load forecasting unit is used to generate load demand forecast results within the next 1 to 24 hours based on the LSTM model.

[0193] 4. The medium- and long-term load forecasting module is used to predict the power load trend in the next few days to weeks based on the data processed by the data preprocessing module, including:

[0194] XGBoost model building unit, used to build XGBoost ensemble learning model to meet the needs of medium- and long-term load forecasting;

[0195] The medium- and long-term load forecasting unit is used to generate load trend forecast results for the next few days to weeks based on the XGBoost model.

[0196] 5. The dispatch decision module is used to generate an optimized dispatch strategy based on the short-term load forecast results and the medium- and long-term load forecast results, including:

[0197] A deep Q network (DQN) model building unit, which is used to build and train a DQN model based on reinforcement learning to generate a scheduling strategy that adapts to the current load demand and resource status;

[0198] A strategy generation unit is used to determine the generation plan of distributed energy, the charging and discharging sequence of energy storage equipment, and the demand response strategy on the load side according to the strategy output by the DQN model;

[0199] The cost optimization unit is used to evaluate the cost of the scheduling strategy and optimize the scheduling scheme based on resource constraints and load balance.

[0200] 6. Resource coordination control module, used to execute the above scheduling optimization strategy and coordinate the operation of distributed energy, energy storage equipment and load-side resources, including:

[0201] Distributed energy control unit, used to dynamically adjust the power generation of distributed energy and perform grid-connection or off-grid control according to load changes;

[0202] The energy storage device control unit is used to execute the charging and discharging strategy of the energy storage device and monitor the status of the device in real time to ensure its safety;

[0203] The load-side control unit is used to control the response of adjustable resources on the load side to achieve demand response and peak shaving.

[0204] 7. Monitoring module, used to monitor the scheduling execution effect in real time, including:

[0205] The status monitoring unit is used to collect the operating status of distributed energy, energy storage equipment and load-side resources in real time;

[0206] Effect analysis unit, used to evaluate the effect of dispatch execution, including deviation analysis from dispatch target, strategy execution cost, adaptability of load response, etc.;

[0207] Alarm unit, used to detect abnormal conditions and trigger alarm signals to ensure the stability of the system.

[0208] 8. Feedback optimization module, which is used to adaptively adjust the parameters of the DQN model based on the feedback data of the monitoring module and update the scheduling optimization strategy, including:

[0209] The parameter adjustment unit is used to adjust the key parameters of the DQN model, such as the learning rate, exploration rate, and discount factor, to improve the model's responsiveness to real-time feedback;

[0210] The experience replay unit is used to store feedback data into the experience replay pool and update the DQN model through training with small batches of samples;

[0211] The policy optimization unit is used to generate new scheduling policies and test them in a simulation environment to ensure the policy adaptability and robustness.

[0212] Wherein: the data acquisition module is connected with the data preprocessing module, and is used to transmit multi-source data to the data preprocessing module for cleaning, outlier removal and normalization;

[0213] The data preprocessing module is connected to the short-term load forecasting module and the medium- and long-term load forecasting module respectively, and is used to transmit the preprocessed data to the short-term load forecasting module and the medium- and long-term load forecasting module to generate short-term load forecasting results and medium- and long-term load forecasting results;

[0214] The output ends of the short-term load forecasting module and the medium- and long-term load forecasting module are connected to the dispatching decision module to provide forecast results to the dispatching decision module so that the dispatching decision module can generate resource allocation and dispatching optimization strategies;

[0215] The dispatch decision module is connected to the resource coordination control module to transmit the generated dispatch optimization strategy to the resource coordination control module to control the operating status of distributed energy, energy storage equipment and load-side resources;

[0216] The resource coordination and control module is connected to the monitoring module to transmit the operating status and scheduling execution status of each resource to the monitoring module to monitor and evaluate the execution effect of the scheduling strategy in real time;

[0217] The monitoring module is connected to the feedback optimization module to transmit the monitoring feedback data to the feedback optimization module to adaptively adjust the DQN model parameters and scheduling optimization strategy;

[0218] The feedback optimization module is connected to the scheduling decision module to transmit the adjusted DQN model parameters and the updated scheduling optimization strategy to the scheduling decision module, thereby optimizing the scheduling performance in a dynamic environment.

[0219] An electronic device or medium, on which a computer program is mounted or stored, wherein: when the above-mentioned computer program is executed by a processor, the above-mentioned virtual power plant peak-shaving optimization scheduling method is implemented.

[0220] The beneficial effects of the present invention are:

[0221] 1. Improve load forecasting accuracy

[0222] The present invention uses the short-term load forecasting and medium- and long-term load forecasting modules, respectively applying the long short-term memory (LSTM) neural network and the XGBoost integrated learning model, to achieve more accurate forecasts for short-term and medium- and long-term load demand changes. Short-term forecasts are used for real-time scheduling decisions, and medium- and long-term forecasts are used for trend analysis and long-term planning, thereby effectively improving the accuracy and reliability of the forecast.

[0223] 2. Realize real-time dynamic scheduling optimization

[0224] The present invention introduces a reinforcement learning algorithm based on a deep Q network (DQN) to generate an optimized resource scheduling strategy. Through continuous learning and real-time updating, the DQN algorithm can automatically generate a dynamic scheduling scheme based on load forecast results and multiple constraints, balance scheduling costs and resource utilization efficiency, and meet the dynamic response requirements of virtual power plants.

[0225] 3. Enhance the collaborative control capabilities of resources

[0226] Through the resource coordination and control module, the present invention can flexibly adjust the operating status of distributed energy, energy storage equipment and load-side resources according to the scheduling optimization strategy, realize efficient coordination and dynamic management of multiple resources, effectively meet the peak-shaving needs of virtual power plants, and improve overall scheduling efficiency.

[0227] 4. Real-time monitoring and adaptive feedback optimization

[0228] The monitoring module and feedback optimization module of the present invention enable the system to monitor the scheduling execution effect in real time, and automatically adjust the DQN model parameters and update the scheduling strategy according to the monitoring feedback data. Through continuous adaptive optimization, the system can quickly respond to environmental changes and load fluctuations, ensure the efficiency and adaptability of the scheduling plan, and thus continuously optimize system performance.

[0229] 5. Reduce operating costs and improve system stability

[0230] The scheduling decision module of the present invention can maximize resource utilization efficiency and reduce overall operating costs when allocating resources based on the cost optimization algorithm. In addition, through real-time monitoring and feedback adjustment of scheduling execution effects, the present invention can quickly adjust strategies in abnormal situations to improve the stability and reliability of the system.

[0231] 6. Support the efficient use of distributed energy and peak load shifting

[0232] The present invention achieves the goals of peak load shaving and valley filling of virtual power plants and demand response through refined management and coordinated scheduling of distributed energy. Through accurate load forecasting and optimized scheduling strategies, the utilization rate of distributed energy and energy storage equipment can be improved, further reducing dependence on traditional power grids and helping to build a low-carbon and efficient energy system.

[0233] In summary, the present invention provides an innovative virtual power plant peak-shaving optimization scheduling method and system, which can not only accurately predict load and dynamically optimize resource allocation, but also continuously optimize scheduling strategies through real-time feedback, thereby significantly improving the operating efficiency and economy of the virtual power plant, and providing an effective solution for achieving efficient utilization of a high proportion of distributed energy, and has good application prospects and economic benefits. BRIEF DESCRIPTION OF THE DRAWINGS

[0234] Figure 1 : A schematic structural diagram of the present invention; DETAILED DESCRIPTION

[0235] The present invention will be further described below in conjunction with the accompanying drawings and specific embodiments:

[0236] Example 1: Figure 1 As shown, a peak load optimization scheduling method for a virtual power plant is provided. The method realizes dynamic optimization scheduling of a distributed energy system based on multi-source data processing and intelligent algorithms. The specific steps are as follows:

[0237] 1. Data collection and preprocessing

[0238] Collect multi-source data such as power generation, equipment status, and environmental information from distributed energy, energy storage equipment, and load side. Clean the data, remove outliers, and normalize them to ensure data consistency and accuracy.

[0239] 2. Short-term load forecasting

[0240] The long short-term memory (LSTM) neural network model is used to predict the power load demand for the next 1 to 24 hours. The preprocessed data is input into the LSTM model, and the short-term load forecast results are generated through the time series analysis of the model.

[0241] 3. Medium and long-term load forecasting

[0242] Use the XGBoost ensemble learning model to predict the load trend in the next few days to weeks. Input the preprocessed data into the XGBoost model, and use the feature recognition ability of the model training to predict the medium- and long-term load change trend.

[0243] 4. Scheduling decision generation

[0244] Based on the short-term and medium-term load forecast results, the Deep Q Network (DQN) model is used to generate optimized dispatch strategies. The DQN model uses a reinforcement learning algorithm to balance dispatch costs and resource constraints, and dynamically generates dispatch optimization strategies for distributed energy generation, energy storage charging and discharging, and load-side response.

[0245] 5. Resource coordination and control

[0246] The generated scheduling optimization strategy is transmitted to the control module to control the power generation plan of distributed energy, the charging and discharging cycle of energy storage equipment and the response behavior of the load side to achieve the coordinated cooperation of various resources.

[0247] 6. Operation effect monitoring

[0248] The monitoring module collects the effect data of scheduling execution in real time and evaluates the actual operation effect of the optimization strategy, including the deviation between the scheduling strategy and the actual load demand and the resource utilization efficiency.

[0249] 7. Feedback optimization and adjustment

[0250] Based on the monitoring feedback data, the parameters of the DQN model are adaptively adjusted. The automatically adjusted DQN model updates the scheduling optimization strategy, thereby continuously adapting to environmental changes and load fluctuations and continuously optimizing scheduling performance.

[0251] Through the scheduling method of this embodiment, the virtual power plant can effectively utilize distributed energy and energy storage resources, smooth out peaks and fill valleys, improve overall operating efficiency, reduce energy costs, and at the same time improve the stability and flexibility of the power system.

[0252] Here is an example:

[0253] In a certain urban microgrid, the virtual power plant includes photovoltaic power generation units, wind power generation units, energy storage battery groups, and adjustable loads (such as electric vehicle charging stations and building air conditioning systems). The load demand of the microgrid has significant daytime fluctuations, and photovoltaic and wind power generation have strong uncertainty. The goal is to use the scheduling method of the present invention to optimize resource allocation, achieve peak shaving and valley filling, reduce operating costs, and ensure system stability.

[0254] Step 1: Data collection and preprocessing

[0255] Data collection: The system collects real-time power generation data (power, weather conditions, etc.) from photovoltaic and wind power generation units, obtains charging and discharging status and remaining capacity data from energy storage systems, and collects load demand from electric vehicle charging stations and building air conditioning systems. Other external data includes weather forecasts, historical load records, holiday information, etc.

[0256] Data cleaning: Perform outlier detection on the collected data and remove or correct erroneous data points (for example, clear abnormally high or low data caused by sensor failure).

[0257] Outlier removal and normalization: Normalize the remaining data (for example, normalize the power data to the [0, 1] interval) to ensure that subsequent algorithms can handle different types of data and enhance the generalization ability of the model.

[0258] Step 2: Short-term load forecasting

[0259] Input data construction: The most recent load demand, weather forecast data (temperature, wind speed, irradiance, etc.) and historical power load records are input into the LSTM neural network model. The time window of the model is set to the past 48 hours of data to predict load changes in the next 1 to 24 hours.

[0260] Model training and prediction: The LSTM model is trained using input time series data and predicts the load demand at each future time point in real time based on the actual historical load pattern of the urban microgrid. For example, the forecast shows that there will be two peaks in electricity demand at 12 noon and 6 pm.

[0261] Generation of short-term load forecast results: Generate short-term load forecast results for the next 1 to 24 hours and use the results for real-time scheduling decisions.

[0262] Step 3: Medium and long term load forecasting

[0263] Data input construction: Use historical load demand data, long-term weather data, seasonal trends, holiday information, etc. as input data and input them into the XGBoost integrated learning model.

[0264] Model training and prediction: The XGBoost model performs medium- and long-term predictions on the load demand trends in the next few days. The model handles complex time dependencies and nonlinear relationships through the integrated structure of decision trees. For example, the forecast shows that the load demand will increase significantly on Friday night, while the load demand will be lower on weekends.

[0265] Generation of medium- and long-term load forecast results: Output medium- and long-term load forecast results for the next few days to weeks for long-term planning of dispatching strategies.

[0266] Step 4: Scheduling decision generation

[0267] Algorithm input: Data such as short-term and medium-term load forecast results, resource status, energy storage status and real-time load demand are input into the reinforcement learning algorithm based on the deep Q network (DQN).

[0268] DQN model training and decision-making: The DQN algorithm generates an optimized dispatch strategy based on factors such as power load demand, dispatch cost, distributed energy and energy storage status. For example, for two peak demands in the next few hours, the DQN algorithm generates the following decisions:

[0269] During peak hours at noon and in the evening, energy storage batteries are given priority for discharge to balance the load.

[0270] When photovoltaic and wind power generation is insufficient, adjust the operation of the building air-conditioning system to reduce peak load pressure on electricity consumption.

[0271] Scheduling strategy output: Generate short-term optimization strategies for the next 1 to 24 hours, and medium- and long-term optimization strategies for the next few days.

[0272] Step 5: Resource coordination and control

[0273] Execution of dispatching strategies: The system controls photovoltaic power generation, wind power generation, energy storage battery packs, and building air conditioners and electric vehicle charging stations on the load side based on the dispatching optimization strategy.

[0274] During peak hours at noon and in the evening, the power of the energy storage batteries is released first to reduce the impact on the power grid.

[0275] During low-load periods (such as at night), the system arranges energy storage charging to ensure that the battery has sufficient reserve power before the next peak arrives.

[0276] Load-side management: During peak load periods, the system reduces the building's electricity demand by regulating the building's air-conditioning load (for example, appropriately delaying the air-conditioning start-up time) to achieve load peak shaving.

[0277] Multi-resource collaborative control: organically combining distributed energy generation, energy storage and load-side response to enable efficient collaboration between distributed energy and energy storage equipment to achieve rational utilization and efficient scheduling of resources.

[0278] Step 6: Run the effect monitoring

[0279] Data monitoring: The monitoring module collects the dispatch execution data in real time, including actual load data, distributed energy generation output, charging and discharging status of energy storage equipment, and load response.

[0280] Effect analysis: Deviation analysis is performed on the actual load curve and the predicted load curve to evaluate the execution effect of the current scheduling strategy. For example, after the strategy is executed, whether the peak value of the load curve is reduced and whether the load peak-shaving and valley-filling target is achieved.

[0281] Scheduling effect feedback: The effect data is transmitted to the feedback optimization module for subsequent scheduling optimization and adjustment.

[0282] Step 7: Feedback optimization and adjustment

[0283] Model parameter adjustment: Automatically adjust the parameters of the DQN model based on the data fed back by the monitoring module. For example, if the load peak is not effectively reduced, increase the energy storage discharge or strengthen the load-side response.

[0284] Dispatch strategy update: Regenerate the dispatch strategy based on the adjusted model parameters to cope with the load fluctuations in the next cycle. For example, if the actual off-peak time is shorter than predicted, adjust the strategy to complete energy storage charging in a shorter time to ensure that the battery has sufficient capacity during the next peak.

[0285] Continuous optimization: The system continuously loops the above process, continuously collects feedback data and optimizes the DQN model, so that the scheduling strategy changes adaptively, thereby improving the accuracy and reliability of scheduling.

[0286] Through the above examples, it is proved that the virtual power plant peak-shaving optimization scheduling method of the present invention effectively adapts to the complex and dynamic power demand and resource status, achieves the goal of load peak shaving and valley filling, and realizes adaptive optimization of the strategy through real-time feedback, thereby improving the overall operating efficiency and economy of the system.

[0287] Embodiment 2: A peak load optimization scheduling system for a virtual power plant, which is composed of multiple functional modules to collaboratively implement dynamic scheduling and optimization of virtual power plant resources. The specific modules and functions are as follows:

[0288] 1. Data acquisition module

[0289] It includes distributed energy generation data acquisition unit, energy storage equipment status data acquisition unit, load side demand data acquisition unit and environmental data acquisition unit, which respectively collect the real-time operation status, power demand and environmental data of various resources.

[0290] 2. Data preprocessing module

[0291] It includes data cleaning unit, outlier detection and elimination unit and data normalization unit. This module cleans the collected multi-source data, eliminates outliers, and performs normalization to ensure data consistency.

[0292] 3. Short-term load forecasting module

[0293] Contains LSTM model building unit and short-term load forecasting unit. This module uses LSTM model to analyze preprocessed data and predict the power load demand in the next 1 to 24 hours.

[0294] 4. Medium and long-term load forecasting module

[0295] Contains XGBoost model building unit and medium- and long-term load forecasting unit. This module uses XGBoost model to analyze preprocessed data, predict the power load trend in the next few days to weeks, and generate medium- and long-term load forecast results.

[0296] 5. Scheduling decision module

[0297] It includes DQN model building unit, strategy generation unit and cost optimization unit. Based on the short-term and medium-term load forecast results, this module uses the DQN model to generate scheduling optimization strategies, taking into account scheduling costs and resource constraints.

[0298] 6. Resource coordination and control module

[0299] It includes a distributed energy control unit, an energy storage device control unit, and a load side control unit. According to the scheduling optimization strategy, this module dynamically controls the power generation of distributed energy, the charging and discharging sequence of energy storage devices, and the response behavior of the load side to achieve efficient coordination of various resources.

[0300] 7. Monitoring module

[0301] It includes status monitoring unit, effect analysis unit and alarm unit. This module monitors the operation status of each resource, evaluates the scheduling execution effect, and triggers alarms in abnormal situations to ensure the stable operation of the system.

[0302] 8. Feedback Optimization Module

[0303] It includes parameter adjustment unit, experience playback unit and strategy optimization unit. Based on the feedback data from the monitoring module, the feedback optimization module automatically adjusts the DQN model parameters and updates the scheduling optimization strategy to adapt to the dynamic changes of the environment and demand, and continuously optimize the system scheduling performance.

[0304] Through the peak-shaving optimization and scheduling system of this embodiment, the virtual power plant can automatically predict and adjust the resource scheduling strategy, achieve efficient resource utilization and accurate load peak shaving and valley filling, thereby optimizing the reliability and flexibility of power supply.

[0305] Embodiment 3: An electronic device or medium having a computer program mounted or stored thereon, wherein the computer program, when executed by a processor, implements the virtual power plant peak load optimization scheduling method of Embodiment 1.

[0306] Finally, it should be noted that the above is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, it is still possible for those skilled in the art to modify the technical solutions described in the aforementioned embodiments or to make equivalent substitutions for some of the technical features therein. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A virtual power plant peak load optimization scheduling method, characterized in that: The following steps are involved: Data collection: Through sensors, IoT terminals and data interfaces, real-time collection of distributed power generation status, load demand, environmental weather data and market price information in virtual power plants; Data preprocessing: cleaning, outlier removal and normalization of collected multi-source data to ensure data consistency and accuracy; Short-term load forecasting, using the long short-term memory neural network model, makes short-term forecasts of power load demand within the next 1 to 24 hours based on preprocessed data and generates short-term load forecast results; Medium- and long-term load forecasting uses the XGBoost ensemble learning model to make medium- and long-term forecasts of power load trends in the next few days to weeks based on preprocessed data and generate medium- and long-term load forecast results; Real-time dispatch decision-making, based on short-term load forecasting results and medium- and long-term load forecasting results, uses a deep Q-network algorithm based on reinforcement learning to generate resource allocation and dispatch optimization strategies; The optimization strategy comprehensively considers dispatching costs, resource constraints and load balance to dynamically adjust the generation, storage charging and discharging and load regulation of distributed energy; Resource coordination and control, executing the dispatch optimization strategy, dynamically managing distributed energy systems, energy storage equipment and load-side resources through the control module, ensuring efficient coordination among various resources to achieve peak load regulation goals; Operation effect monitoring: collect and analyze the effect data of scheduling execution in real time through the monitoring module to evaluate the actual operation effect of the scheduling optimization strategy; Feedback optimization adjustment, based on monitoring feedback data, automatically adjusts the parameters of the deep Q network model and updates the scheduling optimization strategy to cope with environmental changes and demand fluctuations, thereby continuously optimizing the performance of the scheduling method.

2. A virtual power plant peak load optimization scheduling method according to claim 1, characterized in that: The data collection comprises the following steps: Step 1: Data cleaning, wherein the data cleaning includes: Missing value processing: detect missing values ​​in the data set and process the records containing missing values; the missing value processing methods include: interpolation: use the average value of adjacent data or the trend of historical data to fill the missing values; mean replacement: use the mean of the same data feature to replace the missing values; delete records: when the missing value has little impact on the analysis, directly delete the record; Format standardization: standardize the format and units of the collected data, including standardizing temperature data to degrees Celsius and time format to standard time; Step 2: outlier removal, wherein the outlier removal includes: Statistical methods are used to detect outliers by calculating the mean and standard deviation of the data, and extreme values ​​exceeding 3 times the standard deviation are marked or processed; Adopting logical rule detection, setting rules according to the actual operating conditions and physical limitations of the virtual power plant; including limiting the range of voltage, current or power to be within a reasonable operating range, and values ​​outside the range can be regarded as abnormal values; Machine learning methods: For high-dimensional data sets, machine learning anomaly detection algorithms are applied to identify potential abnormal data points by building models; Step 3: normalization processing, wherein the normalization processing includes: Min-max scaling: For variables with real-world numerical ranges, scale the data to a specific interval; the formula is: Z-Score standardization: For data without a fixed range, the data is standardized with a mean of 0 and a standard deviation of 1; the formula is: Among them, μ is the mean and σ is the standard deviation; Specific normalization processing: For seasonal or periodic data, period normalization is used to map the values ​​to the sinusoidal function range of [-1, 1] to retain the periodic characteristics of the data.

3. A virtual power plant peak load optimization scheduling method according to claim 1, characterized in that: The short-term load forecasting comprises the following steps: Step 1: Data preparation, the data preparation includes: Input data selection: Select key characteristic variables that have an impact on short-term load demand forecasting, including historical load data, weather information, time characteristics, and market price data; Dataset division: The preprocessed data is divided into training set, validation set and test set in a ratio of 80%, 10% and 10% to ensure the generalization ability of the model; Time window division: divide the data into fixed time windows to generate time series input samples to ensure that the model can capture time-dependent features; Step 2: constructing a long short-term memory neural network model, wherein the long short-term memory neural network model construction includes: Model architecture design: Design the number of layers and nodes of the long short-term memory neural network model; including one or more LSTM layers, followed by a fully connected layer to output the predicted value; Hyperparameter setting: Determine the hyperparameters of LSTM, including the number of hidden layer units, time step, learning rate, and batch size; hyperparameters will affect the training effect and prediction accuracy of the model; Loss function selection: The mean square error is used as the loss function to measure the gap between the predicted value and the actual value to optimize the model performance; Step 3: Model training, wherein the model training includes: Forward propagation: input the time series data of the training set into the LSTM model and calculate the predicted output; Back propagation and optimization: Use the gradient descent method to calculate the gradient of the loss function with respect to the model parameters, and gradually adjust the model parameters to minimize the loss function; Validation and early stopping: After each round of training, the validation set data is used to evaluate the model performance and monitor whether overfitting occurs. At the same time, an early stopping strategy is adopted to terminate the training when the validation set loss no longer decreases to prevent overfitting. Step 4: Model testing, the model testing includes: Apply the trained LSTM model to the test set to evaluate its prediction accuracy on unseen data; use the root mean square error and mean absolute error indicators to measure the prediction effect; Compare actual load data with forecast results, analyze the forecast error and performance of the model to determine the feasibility of the model in practical application; Step 5: Generate prediction results, which includes: Short-term forecasting: Based on the test set or real-time input data, the trained LSTM model is used to make short-term forecasts of load demand within the next 1 to 24 hours. Data output: Output the prediction results in time series format for use by subsequent scheduling decision modules; Step 6: Result evaluation and model update, wherein the result evaluation and model update include: Regular evaluation: Regularly evaluate the prediction accuracy of the model based on the actual operation effect of the system, and retrain the model with the latest data to maintain the accuracy of the model; Model parameter adjustment: When the load demand pattern or external environment changes significantly, the adaptability of the forecast can be improved by fine-tuning the model parameters or retraining the model.

4. A virtual power plant peak load optimization scheduling method according to claim 1, characterized in that: The medium- and long-term load forecasting includes the following steps: Step 1: Data preparation, the data preparation includes: Feature selection: Based on the needs of medium- and long-term forecasts, relevant feature variables are selected, including historical load data, weather information, calendar features, and electricity price data; Feature engineering: further feature engineering the data to generate features that help improve prediction accuracy, including adding lagged load features and sliding window mean to enhance the model's perception of load trends; Dataset division: divide the data into training set, validation set and test set; because it is a medium- and long-term prediction, the data set is divided in chronological order to ensure the model's generalization ability for future data; Step 2: XGBoost model construction, wherein the XGBoost model construction includes: Model initialization: Set the basic parameters of the XGBoost model, including the number of trees, tree depth, learning rate, L1 and L2 regularization parameters, to control the complexity of the model and the training speed; Loss function selection: Use mean square error as the loss function to measure the deviation between the model prediction value and the actual value; Step 3: feature importance analysis, the feature importance analysis includes: Before training the model, key features are initially screened through feature importance analysis; XGBoost can rank the importance of each feature and help remove irrelevant or redundant features, thereby simplifying the model and improving training efficiency and prediction accuracy; Step 4: Model training, wherein the model training includes: Iterative training: Input the training set data into the XGBoost model, and build a series of decision trees through the iterative method of gradient boosting. Each tree is used to correct the prediction error of the previous round and gradually improve the overall prediction accuracy. Validation and parameter adjustment: During the training process, the model performance is monitored through the validation set, and the hyperparameters, including tree depth, learning rate, and minimum sample split number, are adjusted using cross-validation or grid search to optimize the model effect and avoid overfitting. Step 5: Model testing, the model testing includes: Apply the trained XGBoost model to the test set data to evaluate its prediction effect on unseen data; use the root mean square error and mean absolute error as indicators to measure the prediction accuracy of the model; Analyze the difference between actual load and predicted load, and further adjust model parameters to ensure the stability of the model in medium- and long-term forecasts; Step 6: Generate prediction results, which includes: Medium- and long-term forecasting: Based on real-time input data from actual applications and the trained XGBoost model, power load is predicted in the next few days to weeks. Data output: Output the prediction results as time series data and provide them to the subsequent scheduling optimization module; Step 7: Result evaluation and model updating, wherein the result evaluation and model updating include: Regular evaluation: During the model application process, the prediction effect is regularly evaluated, and the model performance is analyzed based on the deviation between the medium- and long-term prediction results and the actual load data; Model updating: When load patterns or other external conditions change significantly, retrain the model based on the latest data, or fine-tune model parameters to improve its adaptability to the new data.

5. A virtual power plant peak load optimization scheduling method according to claim 1, characterized in that: The real-time scheduling decision includes the following steps: Step 1: define the scheduling environment, wherein the definition of the scheduling environment includes: State space: Determines the expression of the environmental state, which includes the current generation of distributed energy, the amount of energy storage equipment, load demand, and external grid price information; Action space: defines available dispatching actions, including adjusting the generation of distributed energy, controlling the charge and discharge rate of energy storage equipment, and adjusting the controllable load on the load side; Reward function: Design a reward function to measure the pros and cons of different scheduling schemes; use scheduling cost minimization, energy consumption balance, and resource constraint satisfaction as reward indicators to guide the DQN algorithm to find the optimal strategy; Step 2: DQN model construction, the DQN model construction includes: Network architecture design: Build a deep Q network, including an input layer for receiving state information, multiple hidden layers (fully connected layers or convolutional layers) for extracting features, and an output layer corresponding to the Q values ​​of different scheduling actions; Experience replay mechanism: Set up an experience replay pool to store the information of each executed state, action, reward, and next state, so as to facilitate subsequent small batch random sampling training, break the time correlation of data, and improve learning efficiency; Target network mechanism: Establish a target Q network and regularly copy the parameters of the main network to the target network to stabilize the training process and avoid excessive Q value fluctuations; Step 3: Model training, wherein the model training includes: Initialization: Initialize the system state to the current power system state, including the state of power generation, load and energy storage equipment; Q-value update: At each step, collect the current state and select the optimal action, that is, the action that maximizes the Q-value, calculate the immediate reward according to the reward function, and update the parameters of the deep Q network; The Q value update formula is: Where α is the learning rate, r is the immediate reward, γ is the discount factor, and Q(s′,a′) is the maximum Q value of the next state; Exploration and Utilization: Use the ε-greedy strategy to randomly select a certain proportion of actions, and select the actions corresponding to the current optimal Q value for the rest, in order to improve the algorithm's exploration ability and scheduling effect; Step 4: Real-time decision execution, the real-time decision execution includes: According to the trained DQN model, the Q value of each possible scheduling action is calculated in real time, the action with the maximum Q value is selected, and the resource allocation and scheduling optimization strategy is generated; Convert strategies into control instructions, dynamically adjust the generation, storage charging and discharging rates, and load control of distributed energy, and ensure optimal coordination of various resources under different conditions; Step 5: monitoring the effect of the strategy, including: Continuously monitor the execution effect of the scheduling strategy, record the status, action, reward and final effect of each decision, and store them in the experience replay pool for subsequent optimization and strengthening of the model; Analyze the actual effect and deviation of scheduling. If there is a large deviation, feedback will be given to the training module to help the DQN model to optimize itself. Step 6: strategy update and optimization, the strategy update and optimization includes: Dynamic adjustment: Based on load forecasts, environmental changes, and real-time feedback data, the parameters of the DQN model are continuously adjusted to optimize the accuracy and adaptability of the scheduling strategy; Periodic retraining: As the operation and load patterns of the power system change, the DQN model is regularly retrained using the latest data to ensure that it always meets the scheduling requirements of the current system.

6. A virtual power plant peak load optimization scheduling method according to claim 1, characterized in that: The resource coordination control comprises the following steps: Step 1: receiving a scheduling optimization strategy, wherein the receiving scheduling optimization strategy includes: The control module receives the scheduling optimization strategy generated by the previous DQN, including the generation instructions of distributed energy, the charging and discharging plan of the energy storage equipment, and the response requirements of the load side; Step 2: Distributed energy control, wherein the distributed energy control includes: Power generation regulation: Dynamically adjust the power generation of distributed energy according to the dispatch strategy; distributed power sources include photovoltaic and wind power; when weather conditions or resource fluctuations affect power generation, the control system automatically reduces or increases the power generation to maintain system balance; Grid connection management: When the load demand changes significantly, the control module adjusts the grid connection status of distributed energy to ensure stable power output; and quickly disconnects or isolates some distributed power sources in the event of overload or failure to avoid affecting system safety; Step 3: energy storage device control, the energy storage device control includes: Execution of charging and discharging strategies: According to the scheduling strategy, control the energy storage equipment to charge or discharge in different time periods; arrange charging at low load or when there is excess renewable energy, and discharge at peak load or when the electricity price is high to reduce load pressure; Status monitoring and safety control: Real-time monitoring of the power, temperature and health status of energy storage equipment to ensure that it operates within a safe range; when the device is close to full or the power is low, the charge and discharge rate is automatically adjusted to avoid damage to the device due to overcharging or over-discharging; Step 4: load side control, the load side control includes: Load peak shaving: When the load is at its peak, the adjustable load is controlled to delay or run in different time periods to achieve load peak shaving; for interruptible loads, the load is temporarily shut down or the power is reduced according to the strategy requirements; Demand response: For controllable load resources, through price signals or incentives, guide users to increase electricity consumption during off-peak hours or reduce electricity consumption during peak hours to achieve the effect of demand response; Step 5: Coordination feedback mechanism, the coordination feedback mechanism includes: Feedback data collection: Through sensors and monitoring equipment, the operation data of each resource is collected in real time, including information on power generation, energy storage status, and load changes; the feedback data is transmitted to the scheduling module to ensure real-time monitoring of the execution effect of the scheduling strategy; Abnormality detection and processing: When a large load fluctuation or abnormality of the energy storage device is detected during the control execution process, the dispatching system is promptly notified through the feedback mechanism to regenerate the dispatching strategy or adjust the current control plan to ensure system stability; Step 6: Real-time optimization and adjustment, the real-time optimization and adjustment includes: Dynamically adjust the strategy: During the actual execution process, dynamically fine-tune the control strategy based on feedback data and system operation status; Strategy effectiveness evaluation: Regularly evaluate the execution effect of the control strategy, including the coordinated efficiency of power generation, energy storage and load and the overall load balance of the system; optimize the control model based on the evaluation results to improve the applicability of subsequent scheduling strategies.

7. A virtual power plant peak load optimization scheduling method according to claim 1, characterized in that: The operation effect monitoring comprises the following steps: Step 1: Initialize the monitoring module. The initialization of the monitoring module includes: Parameter configuration: According to the scheduling optimization strategy and system requirements, configure the key monitoring indicators of the monitoring module, including distributed energy generation power, energy storage equipment charging and discharging status, and load response; Monitoring frequency setting: according to the requirements of system dynamic changes, set the collection frequency of various types of data to balance real-time performance and data processing load; Step 2: real-time data collection, the real-time data collection includes: Status data acquisition: Through sensors, smart meters and distributed controllers, real-time status data of distributed energy, energy storage equipment and load side, including voltage, current, power, and charge and discharge rate, is collected; Environmental data collection: Regularly collect environmental factors that affect the scheduling execution effect to ensure that the scheduling strategy adapts to changes in external conditions; Step 3: Data preprocessing, the data preprocessing includes: Data cleaning: Eliminate abnormal values ​​or erroneous data in monitoring data to ensure data accuracy; Data smoothing and filtering: Use moving average or low-pass filtering to smooth out the noise in the data and improve the accuracy of subsequent analysis; Data formatting: Storing or transmitting data in a unified format to provide a standardized data basis for further analysis and evaluation; Step 4: Scheduling effect analysis, the scheduling effect analysis includes: Real-time and planned deviation analysis: Compare monitoring data with the expected scheduling plan and analyze the deviation between actual execution and plan; Key performance indicator evaluation: Based on the key performance indicators defined by the system, calculate the performance of the current scheduling strategy and evaluate whether its effect meets the target; Abnormal detection and alarm: When significant deviations or abnormal conditions are found in the scheduling execution, an alarm is triggered immediately and the abnormality is recorded for subsequent analysis and adjustment; Step 5: Generate an effect evaluation report, which includes: Generate operation effect report: Generate detailed effect evaluation report based on data analysis results; the report includes actual data of scheduling execution, deviation analysis, abnormal records, KPI results, and provides a basis for system decision-making and strategy optimization; Trend analysis and prediction: Based on historical operation data, analyze the changing trends of load demand, power generation, and energy storage utilization indicators to provide reference for future scheduling strategies; Step 6: Feedback and strategy optimization, wherein the feedback and strategy optimization include: Feedback to the scheduling module: Feedback the results of operation effect monitoring to the scheduling optimization module; when significant deviations or continuous anomalies are detected, the scheduling module is advised to adjust the strategy to improve the real-time response capability of the system; Adaptive optimization: Automatically adjust the monitoring frequency and parameter range of the monitoring module according to feedback data to adapt to the real-time demand changes of the system and ensure that the monitoring effect is consistent with the scheduling strategy.

8. A virtual power plant peak load optimization scheduling method according to claim 5, characterized in that: The feedback optimization adjustment comprises the following steps: Step 1: Data feedback collection, the data feedback collection includes: Collect key feedback data: Collect data feedback results from the operation effect monitoring module, including deviation analysis of scheduling execution, abnormal conditions, load fluctuations, and energy storage equipment utilization; Extract model optimization information: Extract indicators related to DQN model performance from feedback data, including load forecast error, dispatch response speed, and peak load regulation effect, to determine whether the model needs to be optimized and adjusted; Step 2: Model performance evaluation, the model performance evaluation includes: Calculate strategy benefits and costs: Calculate the overall benefits and costs of the current scheduling strategy based on the feedback data as an evaluation indicator of the DQN model performance; Determine strategy adaptability: Compare the current scheduling effect with environmental requirements, evaluate the adaptability of the existing DQN model in dealing with load demand fluctuations and environmental changes, and determine whether optimization is needed; Step 3: parameter adjustment and adaptive updating, wherein the parameter adjustment and adaptive updating include: Learning rate adjustment: According to the frequency of environmental changes and demand fluctuations in the feedback, the learning rate of the DQN model is adjusted so that the model can adapt to new data faster or more robustly; Exploration rate adjustment: By increasing or decreasing the exploration rate ε of the ε-greedy strategy, the exploration and utilization of the model can be balanced to improve the adaptability of the model in a dynamic environment; Q-value update parameter optimization: Based on deviation analysis and policy response speed, adjust the discount factor γ of the DQN model to ensure that the model strikes a balance between long-term benefits and short-term responses; Step 4: Strategy retraining, wherein the strategy retraining includes: Experience replay pool update: The latest feedback data is stored in the DQN experience replay pool to ensure that the model learns the latest system dynamic changes; Mini-batch training update: Randomly extract new and old data samples from the experience replay pool, perform mini-batch training, and update the DQN model parameters to adapt them to the current system environment; Target network synchronization: After a training iteration, the parameters of the main network are copied to the target network to stabilize the model update process; Step 5: Update and generate the scheduling strategy, wherein the scheduling strategy update and generation includes: New strategy generation: Use the updated DQN model to generate new scheduling optimization strategies, including resource allocation, energy storage charging and discharging plans, and load-side response measures, to better adapt to current load demand and resource status; Multi-scenario testing and evaluation: Test new strategies in a simulation environment, simulating various environments and demand fluctuations to ensure that the strategy is robust and stable; Step 6: Strategy effect verification and iterative optimization, the strategy effect verification and iterative optimization include: Strategy execution effect monitoring: After applying a new scheduling optimization strategy, continuously monitor its execution effect and collect feedback data; focus on monitoring the effect of the new strategy in load peak reduction, resource coordination and abnormal response; Iterative optimization: When the actual effect does not meet expectations, the next round of parameter adjustment and retraining of the DQN model is automatically triggered to achieve a feedback loop and continuously optimize the performance of the scheduling strategy.

9. A virtual power plant peak load optimization scheduling system according to any one of claims 1 to 8, characterized in that: include: Data acquisition module, used to collect multi-source data of virtual power plants; The data acquisition module includes: a distributed energy generation data acquisition unit, which is used to collect the operating data of the power generation power, voltage and current of each distributed energy in real time; an energy storage device status data acquisition unit, which is used to monitor the information of the power, charge and discharge rate, temperature and health status of the energy storage device; a load side demand data acquisition unit, which is used to obtain the power demand, load fluctuation and demand response status of each load device; an environmental data acquisition unit, which is used to obtain external data of weather information and electricity market prices; The data preprocessing module is used to process the multi-source data collected by the data acquisition module, including: a data cleaning unit, which is used to remove noise data and outliers to ensure data accuracy; an outlier detection and removal unit, which is used to mark and remove extreme data that exceeds the preset range; a data normalization unit, which is used to normalize the data to ensure data consistency; The short-term load forecasting module is used to forecast the short-term load demand in the next 1 to 24 hours based on the data processed by the data preprocessing module, including: an LSTM model building unit, which is used to build and train a long short-term memory neural network model to adapt to the nonlinear and time series characteristics of historical load data; a short-term load forecasting unit, which is used to generate load demand forecast results in the next 1 to 24 hours based on the LSTM model; The medium- and long-term load forecasting module is used to forecast the power load trend in the next few days to weeks based on the data processed by the data preprocessing module, including: an XGBoost model building unit, which is used to build an XGBoost ensemble learning model to meet the needs of medium- and long-term load forecasting; a medium- and long-term load forecasting unit, which is used to generate load trend forecast results in the next few days to weeks based on the XGBoost model; The scheduling decision module is used to generate optimized scheduling strategies based on short-term load forecast results and medium- and long-term load forecast results, including: a deep Q network model building unit, which is used to build and train a DQN model based on reinforcement learning to generate a scheduling strategy that adapts to the current load demand and resource status; a strategy generation unit, which is used to determine the generation plan of distributed energy, the charging and discharging sequence of energy storage equipment, and the demand response strategy on the load side according to the strategy output by the DQN model; a cost optimization unit, which is used to evaluate the cost of the scheduling strategy and optimize the scheduling plan based on resource constraints and load balance; The resource coordination control module is used to execute the scheduling optimization strategy and coordinate the operation of distributed energy, energy storage equipment and load-side resources, including: a distributed energy control unit, which is used to dynamically adjust the power generation of distributed energy and perform grid-connected or off-grid control according to load changes; an energy storage device control unit, which is used to execute the charging and discharging strategy of the energy storage device and monitor the status of the device in real time to ensure its safety; a load-side control unit, which is used to control the response of adjustable resources on the load side to achieve demand response and peak shaving; The monitoring module is used to monitor the scheduling execution effect in real time, including: a status monitoring unit, which is used to collect the operating status of distributed energy, energy storage equipment and load-side resources in real time; an effect analysis unit, which is used to evaluate the scheduling execution effect, including deviation analysis from the scheduling target, strategy execution cost, and adaptability of load response; an alarm unit, which is used to detect abnormal conditions and trigger alarm signals to ensure the stability of the system; The feedback optimization module is used to adaptively adjust the parameters of the DQN model based on the feedback data from the monitoring module and update the scheduling optimization strategy, including: a parameter adjustment unit, which is used to adjust the key parameters of the DQN model, such as the learning rate, exploration rate, and discount factor, to improve the model's responsiveness to real-time feedback; an experience replay unit, which is used to store feedback data in the experience replay pool and update the DQN model through training with small batches of samples; a policy optimization unit, which is used to generate new scheduling strategies and test them in a simulation environment to ensure the adaptability and robustness of the strategies; The data acquisition module is connected to the data preprocessing module and is used to transmit multi-source data to the data preprocessing module for cleaning, outlier removal and normalization; The data preprocessing module is connected to the short-term load forecasting module and the medium- and long-term load forecasting module respectively, and is used to transmit the preprocessed data to the short-term load forecasting module and the medium- and long-term load forecasting module to generate short-term load forecasting results and medium- and long-term load forecasting results; The output ends of the short-term load forecasting module and the medium- and long-term load forecasting module are both connected to the dispatching decision module to provide the forecasting results to the dispatching decision module so that the dispatching decision module can generate resource allocation and dispatching optimization strategies; The scheduling decision module is connected to the resource coordination control module and is used to transmit the generated scheduling optimization strategy to the resource coordination control module to control the operating status of distributed energy, energy storage equipment and load-side resources; The resource coordination control module is connected to the monitoring module and is used to transmit the operating status and scheduling execution status of each resource to the monitoring module to monitor and evaluate the execution effect of the scheduling strategy in real time; The monitoring module is connected to the feedback optimization module and is used to transmit monitoring feedback data to the feedback optimization module to adaptively adjust the DQN model parameters and scheduling optimization strategy; The feedback optimization module is connected to the scheduling decision module and is used to transmit the adjusted DQN model parameters and the updated scheduling optimization strategy to the scheduling decision module, thereby optimizing the scheduling performance in a dynamic environment.

10. An electronic device or medium having a computer program mounted or stored thereon, characterized in that: When the computer program is executed by a processor, the virtual power plant peak load optimization scheduling method according to any one of claims 1 to 8 is implemented.

Images