Multi-microgrid cooperative electric energy scheduling method
Through the method of multi-agent deep reinforcement learning and spatiotemporal feature fusion, the computational complexity and model deviation problems in multi-load power scheduling are solved, efficient and accurate power scheduling is achieved, line loss costs are reduced, and energy utilization is improved.
Patent Information
- Application Number
- CN202510980214.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-07-16
- Publication Date
- 2025-10-21
AI Technical Summary
Existing multi-load power scheduling methods have high computational complexity when dealing with nonlinear constraints and high-dimensional uncertainties, making it difficult to achieve global optimality in large-scale systems, and there are problems of model bias and data dependence.
By adopting the method of multi-agent deep reinforcement learning and spatiotemporal feature fusion, a dynamic weight fusion model is constructed through time series feature expansion, bidirectional LSTM load forecasting and MADDPG power collaborative scheduling to achieve efficient power scheduling.
Significantly reduce line loss costs, improve energy utilization, provide economic-low-carbon collaborative optimization, adapt to complex weather change scenarios, and improve the efficiency and accuracy of power scheduling.
Smart Images
Figure CN120822772A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of energy scheduling, and in particular to a multi-microgrid coordinated power scheduling method. Background Art
[0002] In recent years, multi-load power dispatch has been a research hotspot in the fields of energy internet and smart grids. Its core goal is to achieve economic efficiency, reliability, and low-carbonization of the power system by coordinating the power consumption of various loads, including industrial, commercial, residential, electric vehicles (EVs), and energy storage systems (ESSs). Existing methods fall into four categories: traditional mathematical programming, game theory, data-driven approaches, and distributed optimization.
[0003] Traditional optimization methods build deterministic scheduling models based on mathematical programming tools. For example, mixed integer programming is used to optimize the start and shutdown schedules of high-energy-consuming industrial equipment, significantly reducing electricity costs in industrial parks. Dynamic programming is also used to design time-of-use pricing strategies for electric vehicle charging stations to balance peak and valley loads on the power grid. The advantages of these methods lie in their clear model structure and high solution efficiency, making them suitable for small-scale deterministic scenarios. However, their limitations lie in their difficulty handling nonlinear constraints and high-dimensional uncertainty, and their computational complexity increases dramatically in large-scale systems.
[0004] Game theory approaches coordinate the interests of the power grid, users, and third-party service providers through models such as the Stackelberg game and the Nash game. In 2021, some scholars proposed a demand response framework based on the Stackelberg game, in which the power grid acts as a leader, issuing electricity price signals, and users, as followers, dynamically adjust their electricity consumption. Others have used the Nash game to implement distributed energy trading in multi-microgrid systems. These methods can effectively reflect the interest game relationships under market mechanisms and are suitable for multi-agent decision-making scenarios. However, the existence of their equilibrium solutions is heavily dependent on assumptions (such as rational user response), and the heterogeneity of actual user behavior can lead to model bias.
[0005] Data-driven approaches (such as deep learning and reinforcement learning) exploit nonlinear patterns in historical data to address complex dynamic environments. A multivariate load forecasting model based on long-short-term memory networks has been developed, significantly improving forecast accuracy. Deep reinforcement learning is also being used to dynamically optimize the real-time charging and discharging strategies of electric vehicles to adapt to the random fluctuations of renewable energy. These approaches offer the advantage of processing high-dimensional, unstructured data and supporting online learning and dynamic adjustments. However, they rely on large amounts of high-quality historical data, resulting in significant cold-start issues, poor model interpretability, and the risk of "black-box" decision-making.
[0006] Distributed optimization methods achieve coordinated load scheduling through a decentralized architecture. In 2022, the alternating direction multiplier method was used to coordinate energy transactions across multiple community microgrids, reducing communication overhead. Another proposed edge computing framework based on multi-agent reinforcement learning supports autonomous decision-making by local nodes. These methods offer significant advantages in protecting user privacy (data localization) and improving system scalability. However, their convergence speed is affected by network topology, and strict global optimality cannot be guaranteed.
[0007] Based on this, this application proposes a multi-microgrid collaborative power scheduling method. Summary of the Invention
[0008] The purpose of the present invention is to provide a multi-microgrid collaborative power scheduling method, which provides multi-microgrid collaborative power scheduling that integrates multi-agent deep reinforcement learning and spatiotemporal features. This method can reduce line loss costs, significantly improve energy utilization, and provide new ideas for the economic-low-carbon collaborative optimization of distributed energy systems.
[0009] The purpose of the present invention can be achieved through the following technical solutions: A multi-microgrid coordinated power dispatching method comprises the following steps: Steps for screening and associating microgrid load with weather indicators: This method first expands the time series features of the original weather data to generate a lagged sequence (temperature_lagged 6h), cumulative indicators (72h cumulative wind speed) and composite indicators (perceived temperature, wet-bulb temperature); then the screening is carried out in two stages: in the first stage, the time-varying Granger causality test is used to identify indicators with significant lagged causality (such as the 2-hour leading effect of temperature on load), and the nonlinear correlation strength is quantified by mutual information entropy; in the second stage, the LightGBM time series prediction model is trained, based on the feature permutation importance ranking, and the marginal contribution of each indicator to load forecasting is analyzed in combination with the SHAP value; finally, a dynamic weight fusion model is constructed.
[0010] The steps for building a dual-shot LSTM load forecasting model are as follows: first, multi-scale time series alignment and normalization are performed on historical loads, meteorological data, and date features to construct a mixed input matrix containing lag terms (1h / 3h / 24h) and cumulative effect features (such as 48-hour rainfall); then, a bidirectional LSTM layer is used to scan the time series data in both directions. The forward layer captures the natural evolution of the load, and the backward layer extracts the reverse dependency pattern. The global time series representation is obtained by splicing the two-way hidden states; then, a temporal attention mechanism is introduced to dynamically focus on key time periods (such as load mutation points) with self-attention weights to suppress noise interference; finally, the fully connected layer is connected to output multi-step prediction results, and the number of LSTM layers is automatically searched based on the Bayesian optimizer to ensure prediction accuracy while reducing the risk of overfitting.
[0011] The model training steps are as follows: First, a graph-structured data set is constructed based on the grid topology. Nodes represent microgrid states (load, stored energy), and edge attributes include geographic electrical parameters such as distance and tie line equivalent impedance. The graph structure is used to encode node characteristics and neighborhood relationships, generating a topology-aware embedding vector as input to the critic network. A hierarchical action space is designed. The high-level strategy selects the optimal trading partner through a reward function, while the low-level strategy generates energy storage charging and discharging power and transaction volume instructions from the actor network. A hierarchical decision-making mechanism is designed to separate power trading strategies from local energy storage control. A centralized training-decentralized execution (CTDE) architecture is employed to gradually transition from independent scheduling of a single microgrid to complex interactions among multiple microgrids, ultimately achieving optimal global resource allocation under safety constraints.
[0012] The beneficial effects of the present invention are as follows: the present invention can deploy appropriate power to each microgrid in advance according to weather changes, and can cope with sudden power surges through real-time power scheduling between microgrids; The present invention builds a weather forecast index system to predict the impact of weather on electricity consumption, and relies on a multi-agent reinforcement learning method to achieve efficient scheduling of electricity, avoiding the problems of inefficiency, inaccuracy and lack of robustness in traditional methods. BRIEF DESCRIPTION OF THE DRAWINGS
[0013] The present invention will be further described below with reference to the accompanying drawings.
[0014] Figure 1 This is a flow chart of the multi-microgrid collaborative power dispatching method based on multi-agent deep reinforcement learning and spatiotemporal feature fusion of the present invention; Figure 2 This is a flow chart for constructing the load weather index system in the present invention; Figure 3 This is a diagram of the double-layer LSTM structure in the present invention; Figure 4 This is the MADDPG architecture diagram of the present invention. DETAILED DESCRIPTION
[0015] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making any creative efforts shall fall within the scope of protection of the present invention.
[0016] See also Figure 1-Figure 4 As shown, the present invention is a multi-microgrid coordinated power scheduling method, comprising the following steps: Construction of load weather index system Based on historical weather information and historical microgrid load information, the original weather data is first expanded with time series features to generate lag series, cumulative indicators, and composite indicators. This is then screened in two stages: In the first stage, the time-varying Granger causality test is used to identify indicators with significant lagged causality, and the strength of nonlinear association is quantified by mutual information entropy; In the second stage, the LightGBM time series forecasting model is trained, and the marginal contribution of each indicator to load forecasting is analyzed based on the importance ranking of feature permutation, and the SHAP value is combined; finally, a dynamic weight fusion model is constructed to give the indicators related to power load in the weather indicators. The specific process is as follows: Figure 2 shown.
[0017] Microgrid load forecasting based on dual LSTM Microgrid load demand is affected by multiple factors such as weather changes, user behavior, and holidays, and exhibits significant nonlinear and periodic fluctuations. Statistical models struggle to adequately model such complex relationships. However, LSTM, through the coordinated operation of forget gates, input gates, and output gates, can adaptively filter key information from historical load data, memorize electricity consumption patterns across time periods (such as diurnal and weekly cycles), and simultaneously integrate external covariates (such as weather data like temperature and humidity) to achieve multimodal feature fusion. This application integrates multi-dimensional time series data including historical load curves, weather forecasts, user demand trends, etc. through a sliding time window, transfers cross-period memory through hidden states, and uses the load usage of the previous two weeks to output the load forecast value of the microgrid for the next day, providing a high-precision decision-making basis for power dispatching; Exemplary: Microgrid load forecasting method based on Bi-layer Long Short-Term Memory Network (BILSTM). LSTM, as a variant of Recurrent Neural Network (RNN), can effectively capture the long-term dependencies and dynamic time series characteristics in microgrid load data with its unique gating mechanism. The framework is as follows: Figure 3 shown.
[0018] BILSTM consists of two LSTMs: one for forward sequence processing and the other for backward sequence processing. After processing, the outputs of the two LSTMs are concatenated, which can better filter prior information and improve prediction accuracy.
[0019] This application is based on the weather indicators obtained in the load weather indicator system construction step. , minimum temperature , average temperature , air humidity , rainfall and the microgrid load in the previous two weeks As input, the load forecast value of the microgrid for the next day is calculated as shown in the following formula: By dynamically updating the prediction through the sliding window, a prediction prior is provided for the load scheduling between microgrids, thus realizing the closed-loop optimization of "prediction-scheduling".
[0020] Coordinated dispatch of power among microgrids based on MADDPG: MADDPG is a reinforcement learning algorithm for multi-agent collaboration and competition scenarios. Its core adopts the "centralized training, decentralized execution" framework. Each agent has an independent Actor network to generate local strategies, while the Critic network evaluates the value of actions based on global information.
[0021] The algorithm effectively solves the non-stationary problem in multi-agent environments by combining experience replay and policy gradient optimization, supports continuous action space decision-making, and is suitable for collaborative tasks between nodes in complex systems. It can not only achieve individual goal optimization, but also promote multi-agent collaboration through the global perspective of Critic, taking into account the overall system efficiency and individual autonomy. The MADDPG architecture diagram is shown below. Figure 4 shown.
[0022] To reduce power loss and save economic costs, thus achieving optimal distribution and regulation of power under a multi-microgrid architecture; The present invention designs the MADDPG model as follows: Status: Single microgrid Status includes current load , predicted load , microgrid Euclidean distance to other microgrids 、 、 、……、 , the equivalent impedance of the tie lines between microgrids 、 、 、……、 , the state vector is shown as follows: , , ,……, , , ,……, Action: The agent's action space includes the selection of microgrids for load conversion ( ) and the amount of power required ( ), Indicates that the microgrid The existing power, Represents the existing power of the microgrid selected for load conversion, and the action vector is shown as follows: Reward: To improve the economy, reliability, and efficiency of power transmission, this method designs a reward function based on these three aspects.
[0023] In terms of economic efficiency, the key considerations are the electricity price, purchase cost, and energy storage loss, as shown in the following formula: = - - | | Indicates the output power and input power, and Indicates the selling price and buying price, Represents the loss weight, which is approximately proportional to the equivalent impedance between nodes. Indicates the Euclidean distance for transporting electricity.
[0024] For reliability, the focus is on voltage stability, and penalties are imposed for voltage instability. Considering that the farther away the node is, the greater the penalty for exceeding the voltage limit is, the reliability reward is designed to be inversely proportional to the distance between nodes, as shown in the following formula: in, , is the distance normalization parameter.
[0025] As for transmission efficiency, in order to reduce power transmission loss, this method uses the transmission distance to define the transmission efficiency reward, as shown in the following formula: Based on the above rewards, the reward function of this method is as follows: in, 、 as well as Indicates the weight of the corresponding reward. Considering that rewards should have different emphases in different situations, the above three weights should have adaptive adjustment capabilities. When the total power consumption is large, the weight should be increased. , to ensure the reliability of power transmission, when the total power dispatch is large, it should be increased , ensuring the transmission efficiency of electricity and reducing waste.
[0026] By executing the above three steps in sequence and relying on the power dispatch between microgrids, user costs and power costs are reduced, and the power dispatch efficiency is improved, solving the problem of dynamic matching of "source-load-storage" in geographically distributed microgrid groups.
[0027] This application can carry out efficient scheduling of electricity for complex weather change scenarios, has very strong universality, and improves the efficiency of electricity scheduling.
[0028] The above is a detailed description of an embodiment of the present invention. However, the content described is only a preferred embodiment of the present invention and should not be considered to limit the scope of the present invention. All equivalent changes and improvements made within the scope of the present invention should still fall within the scope of the patent coverage of the present invention.
Claims
1. A multi-microgrid coordinated power dispatching method, characterized in that: The following steps are involved: Construction of a microgrid load weather indicator system: This system integrates time series feature engineering with a multi-dimensional feature screening mechanism to generate lagged series, cumulative indicators, and composite indicators from raw weather data. Time-varying Granger causality tests are used to identify indicators with significant lagged causality. The LightGBM time series forecasting model is trained, ranked based on feature permutation importance, and analyzed using SHAP values to analyze the marginal contribution of each indicator to load forecasting. Finally, a dynamic weight fusion model is constructed. Construction of a dual-shot LSTM load forecasting model: A bidirectional LSTM layer is used to extract the periodicity and mutation patterns of load data from the forward and backward directions, respectively. A global temporal representation is obtained by concatenating the hidden states of the two paths. A temporal attention mechanism is introduced to dynamically focus on key time periods using self-attention weights to suppress noise interference. Finally, a fully connected layer is connected to output multi-step forecast results. The number of LSTM layers is automatically searched based on the Bayesian optimizer. Coordinated dispatch of electric energy among microgrids based on MADDPG: For the coordinated dispatch scenario of multiple microgrids, by introducing the multi-agent reinforcement learning method MADDPG, a distance-aware compound reward function is constructed, and a hierarchical decision-making mechanism is designed to separate the power trading strategy from the local energy storage control. A centralized training-distributed execution architecture is adopted to gradually transition from independent dispatch of a single microgrid to complex interaction among multiple microgrids, ultimately achieving optimal global resource allocation under safety constraints.
2. A multi-microgrid coordinated power dispatching method according to claim 1, characterized in that: In the process of integrating time series feature engineering and multi-dimensional feature screening mechanism: Lagged weather features are generated through multi-scale sliding windows, and combined with cumulative sunshine duration and perceived temperature derivatives to construct a feature pool with clear physical meaning. The Granger causality test is used to lock the leading weather factors, and the nonlinear contribution of the mutual information entropy and SHAP value evaluation indicators are combined. Finally, the spatiotemporal impact weights of each weather factor are quantified through the dynamic weight fusion model.
3. The multi-microgrid coordinated power dispatching method according to claim 2, characterized in that: The lag series includes a temperature lag of 6 h; Cumulative indicators include 72-hour cumulative wind speed; Composite indicators include perceived temperature and wet-bulb temperature.
4. The multi-microgrid coordinated power dispatching method according to claim 1, characterized in that: A two-layer LSTM algorithm is used to predict the load of each microgrid, including multi-scale time series alignment and normalization of historical load, meteorological data and date features, and the construction of a mixed input matrix containing lag terms and cumulative effect features.
5. A multi-microgrid coordinated power dispatching method according to claim 4, characterized in that: The lag terms include 1h, 3h, and 24h; Cumulative validation characteristics include 48-hour rainfall.
6. A multi-microgrid coordinated power dispatching method according to claim 4, characterized in that: The bidirectional LSTM layer scans the time series data in both directions. The forward layer captures the natural evolution of the load, and the backward layer extracts the reverse dependency pattern. The global time series representation is obtained by splicing the two-way hidden state.
7. The multi-microgrid coordinated power dispatching method according to claim 1, characterized in that: A MADDPG-based hierarchical reinforcement learning method is used to intelligently dispatch power demand on the power grid, including: Graph structure data is constructed based on the power grid topology, with nodes representing the microgrid status and edge attributes including distance and tie line equivalent impedance geographic electrical parameters; Utilize graph structure to encode node features and neighborhood relationships, and generate topology-aware embedding vectors as input to the Critic network. Global topology information is integrated into the centralized critic network, and the decentralized actor network generates energy storage scheduling and trading strategies based on local observations.
8. The multi-microgrid coordinated power dispatching method according to claim 7, characterized in that: The hierarchical decision-making mechanism includes high-level strategies and low-level strategies; The high-level strategy selects the optimal transaction object through the reward function, and the low-level strategy generates energy storage charging and discharging power and transaction volume instructions by the Actor network.
9. The multi-microgrid coordinated power dispatching method according to claim 8, characterized in that: The reward function construction process includes: The designed reward function has two major features: multi-objective coupling, dynamic adaptation, and physical constraint embedding: Multi-dimensional trade-offs: quantify the comprehensive benefits of the scheduling strategy through linear weighting of the three sub-incentives of economy, reliability, and low carbon; Geophysical constraints are explicitly encoded, and transmission losses caused by distance and impedance between microgrids are embedded in the transaction reward calculation, allowing the strategy to autonomously avoid high-loss long-distance transmission behavior.