Electric power energy storage scheduling method and device based on artificial intelligence

Abstract

The invention discloses an electric power energy storage scheduling method and device based on artificial intelligence, and relates to the technical field of electric power scheduling. First source data are formed by obtaining historical wind power and wind speed data of a target wind power plant; preprocessing the wind power characteristic matrix to obtain a wind power characteristic matrix, inputting the wind power characteristic matrix into a wind power prediction model, and obtaining a wind power output sequence of the target wind power plant in a first target time period; acquiring state data of a plurality of energy storage facilities, a power grid electric load demand sequence and a step electricity price to form second source data; and constructing environment state characteristics based on the wind power output sequence and the second source data, and inputting the environment state characteristics into the pre-training agent to obtain the charging and discharging power of each energy storage facility in the second target time period. By introducing wind power prediction, the foresight and global optimization of energy storage scheduling are realized, the wind power utilization rate is improved, the energy waste is reduced, and the economy and stability of micro-grid operation are enhanced.

Description

Technical Field

[0001] This invention relates to the field of power dispatching technology, specifically to a power energy storage dispatching method and device based on artificial intelligence. Background Technology

[0002] With the rapid development of renewable energy, wind power is playing an increasingly important role in the power system. However, due to the significant randomness and intermittency of wind energy resources, its output fluctuations pose a considerable challenge to the safe and stable operation of the power grid. Wind speed is frequently affected by factors such as climate and terrain, making it difficult to accurately predict the output power of wind farms. Traditional prediction methods based on physical or statistical models have limited accuracy in complex environments, which can easily lead to the accumulation of prediction errors and affect the reliability of power grid power balance and dispatch decisions.

[0003] On the other hand, to improve wind power absorption and peak shaving, energy storage technology is gradually being introduced into microgrid systems. However, traditional energy storage dispatching methods mostly employ rule-based control, linear programming, or heuristic algorithms. While these methods can yield feasible solutions in static scenarios, they lack real-time performance and adaptability under the influence of multiple factors such as wind power fluctuations, electricity price changes, and dynamic load adjustments, making it difficult to achieve refined energy management. Simultaneously, the introduction of time-of-use pricing and tiered pricing mechanisms in the electricity market has made the economic constraints of microgrid operation more complex. Dispatch strategies must not only consider system balance and equipment constraints but also maximize revenue. Existing dispatching methods struggle to address multi-objective optimization problems in high-dimensional dynamic environments, leading to issues such as unreasonable energy storage resource dispatching, low wind power utilization, and high grid operating costs. Summary of the Invention

[0004] The purpose of this invention is to solve the problem of unreasonable energy storage resource scheduling mentioned in the background art, and to propose an artificial intelligence-based power energy storage scheduling method and device.

[0005] A first aspect of this invention provides an artificial intelligence-based power storage dispatching method, wherein the target microgrid includes multiple wind farms and multiple energy storage facilities; the method includes: Historical wind power data and historical wind speed data of the target wind farm are obtained and denoted as the first source data; the target wind farm is any one of multiple wind farms. The first source data is preprocessed to obtain a wind power feature matrix, and the wind power feature matrix is ​​input into the corresponding wind power prediction model to obtain the wind power output sequence of the target wind farm in the first target time period; the first target time period includes multiple time steps. The current status data of multiple energy storage facilities, the electricity load demand sequence of the target microgrid in the first target time period, and the tiered electricity price are obtained and recorded as the second source data; Based on the wind power output sequence and the second source data, environmental state characteristics are constructed; The environmental state features are input into a pre-trained agent to obtain the charging and discharging power of each energy storage facility in the second target time period; the second target time period is the first time step of the first target time period.

[0006] Optionally, the wind power prediction model includes a global dependency extraction branch, a local feature extraction branch, a feature fusion layer, and a prediction layer; wherein: The global dependency extraction branch is used to process the wind power feature matrix using a single Transformer encoder to obtain global dependency features; and to perform dimensionality transformation and feature enhancement on the global dependency features through a temporal convolutional network to obtain the first feature to be fused; the input of the Transformer encoder does not introduce position encoding; The local feature extraction branch is used to process the features at each time step in the wind power feature matrix using a fully connected layer to obtain local detail features; and to perform dimensional transformation and feature enhancement on the local detail features through a temporal convolutional network to obtain a second feature to be fused; the second feature to be fused has the same dimension as the first feature to be fused. The feature fusion layer is used to perform weighted fusion of the first feature to be fused and the second feature to be fused using adaptive weights to obtain global-local collaborative features. The prediction layer is used to map the global-local collaborative features into single-step wind power prediction values ​​using a fully connected network.

[0007] Optionally, obtaining the electrical load demand sequence of the target microgrid during the first target time period includes: Historical electricity load demand data of the target microgrid is obtained, and a load characteristic matrix is ​​constructed by combining weather and time information. The load feature matrix is ​​input into a pre-trained load prediction model to obtain the electrical load demand sequence of the target microgrid during the first target time period; the load prediction model is a long short-term memory network model.

[0008] Optionally, the agent's policy network includes an information focusing layer and an action mapping network; wherein: The information focusing layer is used to weight the input environmental state features using a self-attention mechanism to obtain weighted environmental state features. The operation process of the information focusing layer includes: linearly mapping the environmental state features to a query matrix Q, a key matrix K, and a value matrix V through a preset weight matrix; calculating the inner product of matrix Q and matrix K, and normalizing it using the softmax function to obtain attention weights; multiplying the attention weights by matrix V to obtain the weighted environmental state features. The action mapping network is used to map the weighted environmental state features into the charging and discharging power of energy storage using a fully connected network.

[0009] Optionally, the training process of the agent is divided into two stages, specifically including: Phase 1: Acquire historical optimized scheduling data of the target microgrid and extract state-action pair datasets from it as expert knowledge; The policy network of the agent is pre-trained using the mean square error between the charging and discharging actions output by the agent and the charging and discharging actions in the expert knowledge as the loss function, so as to obtain the initial parameters of the policy network. The second stage involves placing the agent in the simulation environment of the target microgrid based on the initial parameters of the policy network obtained through pre-training. Deep reinforcement learning training is then carried out using the Markov decision process framework to iteratively optimize the policy network.

[0010] A second aspect of this invention provides an artificial intelligence-based power storage dispatching device, wherein the target microgrid includes multiple wind farms and multiple energy storage facilities; the device includes: The first source data acquisition module is used to acquire historical wind power data and historical wind speed data of the target wind farm, denoted as the first source data; the target wind farm is any one of multiple wind farms. The power output prediction module is used to preprocess the first source data to obtain a wind power feature matrix, and input the wind power feature matrix into the corresponding wind power prediction model to obtain the wind power output sequence of the target wind farm in the first target time period; the first target time period includes multiple time steps. The second source data acquisition module is used to acquire the current status data of multiple energy storage facilities, the electricity load demand sequence of the target microgrid in the first target time period, and the tiered electricity price, which is referred to as the second source data. The scheduling scheme generation module is used to construct environmental state features based on the wind power output sequence and the second source data; input the environmental state features into a pre-trained agent to obtain the charging and discharging power of each energy storage facility in the second target time period; the second target time period is the first time step of the first target time period.

[0011] Optionally, the wind power prediction model includes a global dependency extraction branch, a local feature extraction branch, a feature fusion layer, and a prediction layer; wherein: The global dependency extraction branch is used to process the wind power feature matrix using a single Transformer encoder to obtain global dependency features; and to perform dimensionality transformation and feature enhancement on the global dependency features through a temporal convolutional network to obtain the first feature to be fused; the input of the Transformer encoder does not introduce position encoding; The local feature extraction branch is used to process the features at each time step in the wind power feature matrix using a fully connected layer to obtain local detail features; and to perform dimensional transformation and feature enhancement on the local detail features through a temporal convolutional network to obtain a second feature to be fused; the second feature to be fused has the same dimension as the first feature to be fused. The feature fusion layer is used to perform weighted fusion of the first feature to be fused and the second feature to be fused using adaptive weights to obtain global-local collaborative features. The prediction layer is used to map the global-local collaborative features into single-step wind power prediction values ​​using a fully connected network.

[0012] Optionally, the second source data acquisition module includes: The load data acquisition module is used to acquire historical electricity load demand data of the target microgrid and construct a load characteristic matrix by combining weather and time information. The load forecasting module is used to input the load feature matrix into a pre-trained load forecasting model to obtain the electrical load demand sequence of the target microgrid in the first target time period; the load forecasting model is a long short-term memory network model.

[0013] Optionally, the agent's policy network includes an information focusing layer and an action mapping network; wherein: The information focusing layer is used to weight the input environmental state features using a self-attention mechanism to obtain weighted environmental state features. The operation process of the information focusing layer includes: linearly mapping the environmental state features to a query matrix Q, a key matrix K, and a value matrix V through a preset weight matrix; calculating the inner product of matrix Q and matrix K, and normalizing it using the softmax function to obtain attention weights; multiplying the attention weights by matrix V to obtain the weighted environmental state features. The action mapping network is used to map the weighted environmental state features into the charging and discharging power of energy storage using a fully connected network.

[0014] Optionally, the device further includes an agent training module for two-stage training of the agent. Specifically, the agent training module includes: The pre-training module is used to acquire historical optimized scheduling data of the target microgrid and extract state-action pair datasets from it as expert knowledge; the mean square error between the charging and discharging actions output by the agent and the charging and discharging actions in the expert knowledge is used as the loss function to pre-train the policy network of the agent to obtain the initial parameters of the policy network. The interactive learning module is used to place the agent in the simulation environment of the target microgrid for interaction based on the initial parameters of the policy network obtained from pre-training. It combines the Markov decision process framework to carry out deep reinforcement learning training and iteratively optimize the policy network.

[0015] The beneficial effects of this invention are: By forecasting wind power output over a future period, a forward-looking decision-making basis can be established, allowing energy storage device control to shift from relying on lagging real-time measurement data to optimization based on future trends. This enables a comprehensive balance between electricity purchase, sales, and storage over a longer timescale, coordinating energy allocation across different time periods. Compared to passive adjustments based on real-time data, trend-based optimization seeks a balance between maximizing long-term benefits and minimizing operational risks, avoiding energy waste or scheduling conflicts caused by short-term decisions. This leads to more precise and efficient energy storage management, improving the economy and stability of wind power integration and the overall operation of microgrids. Attached Figure Description

[0016] Figure 1 A flowchart illustrating an artificial intelligence-based power storage scheduling method provided in an embodiment of the present invention; Figure 2 A network architecture diagram of a wind power prediction model provided in an embodiment of the present invention; Figure 3 This is a structural diagram of an artificial intelligence-based power storage scheduling device provided in an embodiment of the present invention. Detailed Implementation

[0017] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0018] This invention provides an artificial intelligence-based power storage dispatching method. See also... Figure 1 , Figure 1 A flowchart illustrating an artificial intelligence-based power storage scheduling method provided in this embodiment of the invention. The method includes the following steps: S101, Obtain historical wind power data and historical wind speed data of the target wind farm, and denote them as the first source data.

[0019] S102, preprocess the first source data to obtain the wind power feature matrix.

[0020] S103, input the wind power feature matrix into the corresponding wind power prediction model to obtain the wind power output sequence of the target wind farm in the first target time period.

[0021] S104: Obtain the current status data of multiple energy storage facilities, the electricity load demand sequence and tiered electricity price of the target microgrid in the first target period, and record them as the second source data.

[0022] S105, construct environmental state characteristics based on the wind power output sequence of all wind farms and the second source data.

[0023] S106, input the environmental state characteristics into the pre-trained agent to obtain the charging and discharging power of each energy storage facility during the second target period.

[0024] The target microgrid for this method includes multiple wind farms and multiple energy storage facilities, with the target wind farm being any one of the multiple wind farms. The first target time period includes multiple time steps, and the second target time period is the first time step of the first target time period.

[0025] This invention provides an AI-based power storage dispatching method that predicts wind power output over a future period, enabling forward-looking decision-making. This allows energy storage device control to move beyond relying on lagging real-time measurements and instead optimize based on future trends. This approach comprehensively balances electricity purchase, sales, and storage activities over a longer timescale, coordinating energy allocation across different time periods. Compared to passive adjustments based on real-time data, trend-based optimization seeks a balance between maximizing long-term benefits and minimizing operational risks, avoiding energy waste or dispatch conflicts caused by short-term decisions. This results in more precise and efficient energy storage management, improving the economy and stability of wind power integration and the overall operation of microgrids.

[0026] In one implementation, the first target time period can be set to the next 4 hours, with 15-minute intervals, and the second target time period is the next 15 minutes. The prediction window of the first target time period covers typical wind power and load fluctuation cycles, enabling the dispatch system to anticipate future power trends. The second target time period, as a decision interval, ensures the real-time performance and flexible response capability of energy storage control.

[0027] In one embodiment, see Figure 2 , Figure 2 This is a network architecture diagram of a wind power prediction model provided in an embodiment of the present invention. The wind power prediction model includes a global dependency extraction branch, a local feature extraction branch, a feature fusion layer, and a prediction layer. Wherein: Global Dependency Extraction Branch: First, a Transformer encoder is used to process the wind power feature matrix, and a multi-head self-attention mechanism is used to capture the long-term global temporal dependencies of the data to obtain global dependency features. Then, a temporal convolutional network is used to perform dimensionality transformation and feature enhancement on the global dependency features to obtain the first feature to be fused. The Transformer encoder and temporal convolutional networks are well-known neural network structures to those skilled in the art, and will not be elaborated upon here. It should be noted that this embodiment does not introduce positional encoding at the input of the Transformer encoder. The traditional Transformer uses positional encoding because its self-attention mechanism cannot directly perceive the temporal order of the sequence and needs to retain temporal features by adding positional information. In this embodiment, however, the local feature extraction branch directly processes the original temporal data, enhancing local temporal details while preserving the time step dimension, thus effectively compensating for the temporal feature loss that might occur if the Transformer encoder did not use positional encoding. In contrast, adding additional positional encoding would lead to repetition of temporal information or even interference with feature representation, which is detrimental to the overall performance of the model.

[0028] Local Feature Extraction Branch: First, a fully connected layer is used to process the features at each time step of the wind power feature matrix. A linear mapping is used to enhance key features in the original data, yielding local detail features. Then, a temporal convolutional network is used to perform dimensionality transformation and feature enhancement on the local detail features, resulting in the second feature to be fused. Specifically, the second feature to be fused has the same dimension as the first feature to be fused. The two temporal convolutional networks do not share parameters.

[0029] Feature fusion layer: Adaptive weight P is used to weight and fuse the first and second features to be fused, resulting in global-local collaborative features. : During model training, gradient descent is used to dynamically optimize weights, allowing the model to autonomously allocate the contribution of global and local features, thus avoiding manual parameter tuning.

[0030] Prediction layer: A fully connected network is used to map global and local collaborative features into single-step wind power prediction values.

[0031] This model employs a dual-structure approach, combining a global dependency extraction branch using Transformer and a local feature extraction branch using linear transformation. This allows it to simultaneously capture both long-term temporal correlations and short-term fluctuation details of wind power without requiring manual noise reduction. Consequently, it significantly improves the accuracy and stability of wind power output prediction, providing a more reliable basis for subsequent energy storage scheduling.

[0032] In one embodiment, the state data of the energy storage facility includes its State of Charge (SOC), which can be measured in real time. Tiered pricing, including the purchase price and the retail price of electricity, can be determined in advance. The electricity load demand sequence needs to be obtained through ultra-short-term forecasting, specifically: Step 1: Obtain historical electricity load demand data for the target microgrid, and construct a load characteristic matrix by combining weather and time information.

[0033] Step 2: Input the load feature matrix into the pre-trained load prediction model to obtain the electrical load demand sequence of the target microgrid in the first target time period.

[0034] The load forecasting model can be a Long Short-Term Memory (LSTM) network model.

[0035] This embodiment combines historical load data, weather information, and time information to construct a load feature matrix, which can comprehensively depict the patterns of load changes from multiple dimensions, improving the model's ability to perceive and predict load fluctuations. Historical load data reflects the time-series characteristics of electricity consumption behavior, weather information (such as temperature and humidity) reveals the direct impact of climate conditions on electricity demand, and time information (such as hours, days of the week, holidays, and seasons) reflects the periodicity and regularity of user electricity consumption. Integrating these three types of information to form a load feature matrix allows the prediction model to simultaneously capture long-term trends, periodic patterns, and short-term fluctuations. The LSTM model has excellent memory and time-dependent modeling capabilities, effectively capturing the fluctuation patterns of load at different time scales, thereby significantly improving the accuracy and timeliness of load forecasting.

[0036] In one embodiment, an improved policy network is proposed, which enhances sensitivity to high-influence factors and improves decision-making accuracy by introducing a self-attention mechanism. Specifically, the agent's policy network includes an information focusing layer and an action mapping network; wherein: Information Focusing Layer: This layer employs a self-attention mechanism to weight the input environmental state features X, resulting in weighted environmental state features. The computational process of this information focusing layer includes: using a learnable weight matrix... , , The environmental state characteristics are linearly mapped to a query matrix Q, a key matrix K, and a value matrix V. , , ); calculate the inner product of matrix Q and matrix K, and normalize it using the softmax function to obtain the attention weights AttW. The attention weights are multiplied by matrix V to obtain the weighted environmental state features X. att ( ).

[0037] An action mapping network is used to map weighted environmental state features into energy storage charging and discharging power using a fully connected network. The charging and discharging power is between [-1, 1], representing the ratio of the maximum allowable charging and discharging power. A value greater than 0 indicates charging, and a value less than 0 indicates discharging. The true charging and discharging power is obtained through inverse normalization.

[0038] The information focusing layer uses a self-attention mechanism to weight the input environmental state features, enabling the model to automatically learn and highlight features that significantly impact energy storage scheduling results. This achieves focused attention and noise suppression amidst complex and multidimensional input information. Compared to traditional fully connected structures, this mechanism more effectively captures the importance of different features, improving the model's responsiveness to high-impact factors such as wind power fluctuations, electricity price changes, and load dynamics. The action mapping network maps the weighted features to normalized charging and discharging power outputs, enabling fine-grained control in a continuous action space, thereby enhancing the flexibility, stability, and economy of energy storage system scheduling.

[0039] In one embodiment, the training process of the agent is divided into two stages, specifically: Phase 1: Acquire historical optimized scheduling data of the target microgrid (offline scheduling optimization records performed using linear programming and mixed integer programming), and extract state-action pair datasets from them as expert knowledge; use the mean square error between the charging and discharging actions output by the agent and the charging and discharging actions in the expert knowledge as the loss function to pre-train the agent's policy network and obtain the initial parameters of the policy network.

[0040] The second stage involves placing the agent in the simulation environment of the target microgrid based on the initial parameters of the policy network obtained through pre-training. Deep reinforcement learning training is then carried out using the Markov decision process framework to iteratively optimize the policy network.

[0041] This embodiment significantly improves the learning efficiency, stability, and decision-making performance of the agent by employing a two-stage training mechanism of expert knowledge pre-training and reinforcement learning self-optimization. In the first stage, historical optimization scheduling data is used as expert knowledge to perform supervised pre-training on the policy network, enabling the agent to have reasonable charging and discharging decision-making capabilities from the initial stage, avoiding the training instability and slow convergence problems caused by reinforcement learning's exploration from scratch. In the second stage, based on the pre-trained parameters, the agent continuously interacts in the simulation environment through reinforcement learning, enabling it to adaptively optimize the policy in dynamic environments and further improve its ability to cope with different scenarios.

[0042] In one implementation, the reward function during the second stage of interactive learning can be: ;in: It is an economic incentive used to minimize operating costs. It can be designed as follows: ; It is the revenue from electricity sales; It is the cost of purchasing electricity; This refers to the operating cost of energy storage equipment.

[0043] It is a safety penalty used to ensure power grid stability. It can be designed as follows: ; It is the demand for electrical load; It is the total output of wind power; It is the total charging and discharging power of energy storage.

[0044] This is an operational efficiency reward used to optimize equipment usage. It can be designed as follows: std represents the variance of the State of Charge (SOC) of each energy storage device.

[0045] This is a smoothness-of-operation reward used to improve equipment lifespan. It can be designed as: ; This represents the charging and discharging power of the i-th energy storage device, which the agent decides to execute in the next time period, at the current time t. This represents the charging and discharging power of the i-th energy storage device currently in operation during the current time period.

[0046] , , , These are weighting coefficients, which can be flexibly set by technical personnel. For example, if economy is the primary objective, then... Set it to a larger number.

[0047] This invention provides an artificial intelligence-based power storage dispatching device. See also... Figure 3 , Figure 3 This is a structural diagram of an artificial intelligence-based power storage dispatching device provided in an embodiment of the present invention. The device includes: The first source data acquisition module is used to acquire historical wind power data and historical wind speed data of the target wind farm, which is referred to as the first source data.

[0048] The power output prediction module is used to preprocess the first source data to obtain the wind power feature matrix, and input the wind power feature matrix into the corresponding wind power prediction model to obtain the wind power output sequence of the target wind farm in the first target time period.

[0049] The second source data acquisition module is used to acquire the current status data of multiple energy storage facilities, the electricity load demand sequence of the target microgrid in the first target period, and the tiered electricity price, which is referred to as the second source data.

[0050] The scheduling scheme generation module is used to construct environmental feature vectors based on the wind power output sequence of all wind farms and the second source data; the environmental feature vectors are input into the pre-trained agent to obtain the charging and discharging power of each energy storage facility in the second target time period.

[0051] The target microgrid for this device includes multiple wind farms and multiple energy storage facilities, with the target wind farm being any one of the multiple wind farms. The first target time period includes multiple time steps, and the second target time period is the first time step of the first target time period.

[0052] This invention provides an AI-based power storage dispatching device that predicts wind power output over a future period, enabling forward-looking decision-making. This allows the control of energy storage devices to move beyond relying on lagging real-time measurements and instead optimize based on future trends. This allows for a more comprehensive balance between electricity purchase, sales, and storage over a longer timescale, coordinating energy allocation across different time periods. Compared to passive adjustments based on real-time data, trend-based optimization seeks to maximize long-term benefits while minimizing operational risks, avoiding energy waste or dispatch conflicts caused by short-term decisions. This results in more precise and efficient energy storage management, improving the economy and stability of wind power integration and the overall operation of microgrids.

[0053] In one embodiment, the second source data acquisition module includes: The load data acquisition module is used to acquire historical electricity load demand data of the target microgrid and construct a load characteristic matrix by combining weather and time characteristics.

[0054] The load forecasting module is used to input the load feature matrix into the pre-trained load forecasting model to obtain the electrical load demand sequence of the target microgrid in the first target time period.

[0055] In one embodiment, the device further includes an agent training module for two-stage training of the agent. Specifically, the agent training module includes: The pre-training module is used to acquire historical optimized scheduling data of the target microgrid and extract state-action pair datasets from it as expert knowledge. The mean square error between the charging and discharging actions output by the agent and the charging and discharging actions in the expert knowledge is used as the loss function to pre-train the agent's policy network and obtain the initial parameters of the policy network.

[0056] The interactive learning module is used to place the agent in the simulation environment of the target microgrid for interaction based on the initial parameters of the policy network obtained from pre-training. It combines the Markov decision process framework to carry out deep reinforcement learning training and iteratively optimize the policy network.

[0057] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall fall within the scope of the present invention.

Claims

1. A power storage dispatching method based on artificial intelligence, characterized in that, The target microgrid includes multiple wind farms and multiple energy storage facilities; the method includes: Historical wind power data and historical wind speed data of the target wind farm are obtained and denoted as the first source data; the target wind farm is any one of multiple wind farms. The first source data is preprocessed to obtain a wind power feature matrix, and the wind power feature matrix is ​​input into the corresponding wind power prediction model to obtain the wind power output sequence of the target wind farm in the first target time period; the first target time period includes multiple time steps. The current status data of multiple energy storage facilities, the electricity load demand sequence of the target microgrid in the first target time period, and the tiered electricity price are obtained and recorded as the second source data; Based on the wind power output sequence and the second source data, environmental state characteristics are constructed; The environmental state features are input into a pre-trained agent to obtain the charging and discharging power of each energy storage facility in the second target time period; the second target time period is the first time step of the first target time period.

2. The power storage dispatching method based on artificial intelligence according to claim 1, characterized in that, The wind power prediction model includes a global dependency extraction branch, a local feature extraction branch, a feature fusion layer, and a prediction layer; wherein: The global dependency extraction branch is used to process the wind power feature matrix using a single Transformer encoder to obtain global dependency features; and to perform dimensionality transformation and feature enhancement on the global dependency features through a temporal convolutional network to obtain the first feature to be fused; the input of the Transformer encoder does not introduce position encoding; The local feature extraction branch is used to process the features at each time step in the wind power feature matrix using a fully connected layer to obtain local detail features; and to perform dimensional transformation and feature enhancement on the local detail features through a temporal convolutional network to obtain a second feature to be fused; the second feature to be fused has the same dimension as the first feature to be fused. The feature fusion layer is used to perform weighted fusion of the first feature to be fused and the second feature to be fused using adaptive weights to obtain global-local collaborative features. The prediction layer is used to map the global-local collaborative features into single-step wind power prediction values ​​using a fully connected network.

3. The power storage dispatching method based on artificial intelligence according to claim 1, characterized in that, Obtaining the electrical load demand sequence of the target microgrid during the first target time period includes: Historical electricity load demand data of the target microgrid is obtained, and a load characteristic matrix is ​​constructed by combining weather and time information. The load feature matrix is ​​input into a pre-trained load prediction model to obtain the electrical load demand sequence of the target microgrid during the first target time period; the load prediction model is a long short-term memory network model.

4. The power storage dispatching method based on artificial intelligence according to claim 1, characterized in that, The agent's policy network includes an information focusing layer and an action mapping network; wherein: The information focusing layer is used to weight the input environmental state features using a self-attention mechanism to obtain weighted environmental state features. The operation process of the information focusing layer includes: linearly mapping the environmental state features to a query matrix Q, a key matrix K, and a value matrix V through a preset weight matrix; calculating the inner product of matrix Q and matrix K, and normalizing it using the softmax function to obtain attention weights; multiplying the attention weights by matrix V to obtain the weighted environmental state features. The action mapping network is used to map the weighted environmental state features into the charging and discharging power of energy storage using a fully connected network.

5. The power storage dispatching method based on artificial intelligence according to claim 4, characterized in that, The training process of the intelligent agent is divided into two stages, specifically including: Phase 1: Acquire historical optimized scheduling data of the target microgrid and extract state-action pair datasets from it as expert knowledge; The policy network of the agent is pre-trained using the mean square error between the charging and discharging actions output by the agent and the charging and discharging actions in the expert knowledge as the loss function, so as to obtain the initial parameters of the policy network. The second stage involves placing the agent in the simulation environment of the target microgrid based on the initial parameters of the policy network obtained through pre-training. Deep reinforcement learning training is then carried out using the Markov decision process framework to iteratively optimize the policy network.

6. A power storage dispatching device based on artificial intelligence, characterized in that, The target microgrid includes multiple wind farms and multiple energy storage facilities; the device includes: The first source data acquisition module is used to acquire historical wind power data and historical wind speed data of the target wind farm, denoted as the first source data; the target wind farm is any one of multiple wind farms. The power output prediction module is used to preprocess the first source data to obtain a wind power feature matrix, and input the wind power feature matrix into the corresponding wind power prediction model to obtain the wind power output sequence of the target wind farm in the first target time period; the first target time period includes multiple time steps. The second source data acquisition module is used to acquire the current status data of multiple energy storage facilities, the electricity load demand sequence of the target microgrid in the first target time period, and the tiered electricity price, which is referred to as the second source data. The scheduling scheme generation module is used to construct environmental state features based on the wind power output sequence and the second source data; input the environmental state features into a pre-trained agent to obtain the charging and discharging power of each energy storage facility in the second target time period; the second target time period is the first time step of the first target time period.

7. The power storage dispatching device based on artificial intelligence according to claim 6, characterized in that, The wind power prediction model includes a global dependency extraction branch, a local feature extraction branch, a feature fusion layer, and a prediction layer; wherein: The global dependency extraction branch is used to process the wind power feature matrix using a single Transformer encoder to obtain global dependency features; and to perform dimensionality transformation and feature enhancement on the global dependency features through a temporal convolutional network to obtain the first feature to be fused; the input of the Transformer encoder does not introduce position encoding; The local feature extraction branch is used to process the features at each time step in the wind power feature matrix using a fully connected layer to obtain local detail features; and to perform dimensional transformation and feature enhancement on the local detail features through a temporal convolutional network to obtain a second feature to be fused; the second feature to be fused has the same dimension as the first feature to be fused. The feature fusion layer is used to perform weighted fusion of the first feature to be fused and the second feature to be fused using adaptive weights to obtain global-local collaborative features. The prediction layer is used to map the global-local collaborative features into single-step wind power prediction values ​​using a fully connected network.

8. The power storage dispatching device based on artificial intelligence according to claim 6, characterized in that, The second source data acquisition module includes: The load data acquisition module is used to acquire historical electricity load demand data of the target microgrid and construct a load characteristic matrix by combining weather and time information. The load forecasting module is used to input the load feature matrix into a pre-trained load forecasting model to obtain the electrical load demand sequence of the target microgrid in the first target time period; the load forecasting model is a long short-term memory network model.

9. A power storage dispatching device based on artificial intelligence according to claim 6, characterized in that, The agent's policy network includes an information focusing layer and an action mapping network; wherein: The information focusing layer is used to weight the input environmental state features using a self-attention mechanism to obtain weighted environmental state features. The operation process of the information focusing layer includes: linearly mapping the environmental state features to a query matrix Q, a key matrix K, and a value matrix V through a preset weight matrix; calculating the inner product of matrix Q and matrix K, and normalizing it using the softmax function to obtain attention weights; multiplying the attention weights by matrix V to obtain the weighted environmental state features. The action mapping network is used to map the weighted environmental state features into the charging and discharging power of energy storage using a fully connected network.

10. A power storage dispatching device based on artificial intelligence according to claim 9, characterized in that, The device further includes an agent training module for two-stage training of the agent. Specifically, the agent training module includes: The pre-training module is used to acquire historical optimized scheduling data of the target microgrid and extract state-action pair datasets from it as expert knowledge; the mean square error between the charging and discharging actions output by the agent and the charging and discharging actions in the expert knowledge is used as the loss function to pre-train the policy network of the agent to obtain the initial parameters of the policy network. The interactive learning module is used to place the agent in the simulation environment of the target microgrid for interaction based on the initial parameters of the policy network obtained from pre-training. It combines the Markov decision process framework to carry out deep reinforcement learning training and iteratively optimize the policy network.

Images