A virtual power plant multi-agent collaborative regulation method and system
By combining multi-scale load forecasting and chaotic-deep learning-coupled electricity price forecasting with a three-level multi-agent architecture, the problems of collaborative efficiency and real-time performance of virtual power plants are solved, achieving high-precision load and electricity price forecasting and improving the control efficiency and profitability of virtual power plants.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CCCC HUAZHONG INVESTMENT CO LTD
- Filing Date
- 2026-02-12
- Publication Date
- 2026-06-05
AI Technical Summary
The multi-agent collaborative efficiency and real-time performance of existing virtual power plants need to be improved. The adaptive fusion mechanism of load forecasting across multiple time scales is imperfect. The depth of exploration of the chaotic dynamic characteristics of the sequence in electricity price forecasting is insufficient. There is a lack of deep integration of forecasting, regulation and revenue.
A multi-scale load forecasting model is adopted, combining LSTM and GRU models. Load forecasting is performed through dynamic fusion of entropy weights, and electricity price forecasting is performed by combining chaos theory and deep learning. A three-level multi-agent collaborative architecture of regional level, station level, and equipment level is constructed to achieve optimized resource scheduling.
It improves the accuracy of load forecasting, reduces the deviation in electricity price forecasting, enhances the regulation efficiency and profitability of virtual power plants, meets the real-time needs of market transactions, and achieves coordinated regulation of global optimization and local autonomy.
Smart Images
Figure CN122155198A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of power control technology, specifically relating to a method and system for multi-agent collaborative control of a virtual power plant. Background Technology
[0002] Guided by the "dual-carbon" strategic goal, building a new power system with new energy sources as the mainstay has become the core task of my country's energy revolution. In this process, virtual power plants, as a key technology for aggregating and optimizing distributed energy resources, have shown great potential in improving grid flexibility and promoting the efficient consumption of new energy.
[0003] In terms of control architecture, early approaches focused on centralized optimization methods, which, while achieving global optimum, suffered from computational complexity, high communication pressure, and poor fault tolerance. Subsequently, distributed optimization (such as ADMM) and multi-agent systems (MAS) gradually became mainstream. The latter, by delegating decision-making power to autonomous agents, is more suitable for describing heterogeneous and distributed resource entities in VPP, but its coordination mechanism and communication efficiency still need improvement.
[0004] In predictive modeling, load and electricity price forecasting are crucial foundations for VPPs' participation in market decision-making. Traditional methods such as ARIMA and SVM have limited capabilities in modeling nonlinear and non-stationary sequences. In recent years, deep learning models such as LSTM, GRU, convolutional neural networks (CNN), and attention mechanism models (such as Transformer) have demonstrated outstanding performance in time series forecasting, effectively capturing long-term dependencies and complex patterns. Especially with the rise of large-scale pre-trained models, methods based on transfer learning and domain adaptation have provided new approaches for cross-scenario and cross-regional forecasting tasks. Meanwhile, electricity price sequences, influenced by market mechanisms, policies, weather, and other factors, exhibit significant chaotic characteristics. Combining chaos theory with deep learning for phase space reconstruction and forecasting has become a cutting-edge direction for improving the accuracy of electricity price forecasting.
[0005] In summary, while existing research has made significant progress in VPP architecture design and prediction models, the following shortcomings remain: the efficiency and real-time performance of multi-agent collaboration need improvement; the adaptive fusion mechanism for load forecasting across multiple time scales is still imperfect; and the depth of exploration into the chaotic dynamics of electricity price forecasting is insufficient. Furthermore, existing research often focuses on single technical aspects, lacking a deep integration of "prediction-regulation-revenue." Therefore, a systematic deep learning-driven regulation scheme is urgently needed to fill these research gaps. Summary of the Invention
[0006] The present invention aims to at least partially solve one of the technical problems in the aforementioned related technologies.
[0007] Therefore, the purpose of this invention is to provide a method and system for multi-agent collaborative control of a virtual power plant, which can achieve both global optimization and ensure local autonomy in collaborative control.
[0008] To solve the above-mentioned technical problems, the present invention is implemented as follows: This invention provides a method for multi-agent collaborative control of a virtual power plant, the method comprising: S1. Perform multi-scale load forecasting and output the forecast results; S2. Perform chaotic-deep learning coupled electricity price prediction and output the prediction results; S3. Optimize resource scheduling based on the prediction results of S1 and S2.
[0009] In addition, the virtual power plant multi-agent collaborative control method according to the present invention may also have the following additional technical features: In some implementations, step S1 includes: S11. Collect data and perform data preprocessing to provide input for the split-scale model; S12. Each scaled model makes its own predictions; S13. The results of each prediction are dynamically fused using entropy weights to obtain the results of multi-scale load prediction.
[0010] In some implementations, the scaling model includes an LSTM sub-model, a GRU sub-model, and an entropy weight dynamic fusion model, which respectively fuse long-term prediction, short-term prediction, and long-short-term prediction results.
[0011] In some implementations, the entropy weight dynamic fusion includes: calculating the prediction errors of the LSTM sub-model and the GRU sub-model, determining their weights using the entropy weight method, and obtaining the fused prediction result based on the weights and prediction results as the result of multi-scale load prediction, so as to reduce the error of a single model.
[0012] In some implementations, the data collected in S11 includes multi-dimensional feature data such as historical load data, meteorological data, policy data, holidays, and traffic flow within a specific time period.
[0013] In some implementations, the long-term forecast is: using an LSTM sub-model to capture long-term load dependence patterns and output the load for the next day to the following week; The short-term forecast is: using the GRU sub-model to quickly respond to short-term load fluctuations and output the load for the next 1-6 hours; The fusion of long-term and short-term forecast results is as follows: based on their respective weights, the results of long-term and short-term forecasts are merged to output the load for the next 15-60 minutes.
[0014] In some implementations, step S2 includes: S21. Chaotic Characteristics Identification: The CC method is used to analyze the electricity price time series, calculate the delay time and embedding dimension, and verify the chaotic characteristics of the electricity price series; S22. Phase Space Reconstruction: Mapping a one-dimensional electricity price time series to a high-dimensional phase space, constructing a phase space vector, and revealing the nonlinear dynamic evolution law of electricity prices; S23.LSTM Model Prediction: Using the reconstructed phase space vector as input, the LSTM model is used for prediction. The prediction uses the mean squared error as the loss function and outputs the 15-minute electricity price for the next 1-7 days.
[0015] In some implementations, step S3 includes: S31. Architecture Initialization: Build a three-level intelligent agent architecture of region level, site level, and equipment level, and clarify the responsibilities and communication mechanisms of intelligent agents at each level; S32. Objective function and constraint setting: With the maximization of market revenue as the core objective, resource constraints such as upper and lower limits of regulation capacity, grid constraints such as total grid-connected power limit, response time constraints, and energy storage constraints involving the state of charge range are set. The market revenue maximization mentioned above includes demand response revenue, ancillary service revenue, and spot arbitrage revenue, minus operating costs; S33. Closed-loop coordinated scheduling: A regional-level intelligent agent receives load and electricity price forecasts and formulates global market trading strategies and cross-regional resource coordination plans. Station-level intelligent agent: Receives instructions from regional-level intelligent agents, optimizes resource scheduling within a single station, and provides feedback on execution status; Equipment-level intelligent agent: Executes site-level intelligent agent commands to achieve real-time control of DER equipment, including photovoltaic, energy storage, and charging piles, and provides synchronous status feedback; Dynamic iteration: Adjusting scheduling instructions based on real-time feedback information to form a closed loop of prediction-control-feedback-optimization.
[0016] This invention also provides a virtual power plant multi-agent collaborative control system, capable of implementing the virtual power plant multi-agent collaborative control method as described in any of the preceding embodiments; the system includes: Multi-scale load forecasting module: It is configured to achieve long-term forecasting through the LSTM submodule, short-term forecasting through the GRU submodule, and dynamically calculate weights through the entropy weight method to fuse the long-term and short-term forecasting results and output the load forecasting results. The chaos-deep learning coupled electricity price prediction module is configured to perform chaotic characteristic identification, phase space reconstruction, and LSTM prediction, thereby outputting multi-period electricity price prediction results; and, The three-level multi-agent collaborative control module is configured to construct objective functions and constraint systems, enabling information interaction, collaborative control, and optimization iteration involving regional-level agents, station-level agents, and equipment-level agents.
[0017] In addition, the virtual power plant multi-agent collaborative control system according to the present invention may also have the following additional technical features: In some implementations, the chaos-deep learning coupled electricity price prediction module includes: Chaotic characteristic unit: The delay time and embedding dimension are calculated using the CC method; Phase space reconstruction unit: maps a one-dimensional electricity price sequence into a high-dimensional vector; LSTM prediction model: The reconstructed vector is used as the input of the LSTM model, and the output is the electricity price prediction result for multiple time periods; Error assessment unit: Monitors prediction accuracy in real time using MAPE and RMSE metrics.
[0018] Compared with the prior art, the present invention has at least the following beneficial effects: In this embodiment of the invention, the provided multi-agent collaborative control method for virtual power plants uses LSTM to capture long-term load trends and GRU to quickly respond to short-term fluctuations. After fusion, it can accurately adapt to the scheduling needs of multiple scenarios in virtual power plants. The accuracy is improved by 14.0%-25.0% compared with a single LSTM / GRU model and by more than 40% compared with the traditional ARIMA model, providing highly reliable load demand input for subsequent control decisions. In this embodiment of the invention, the provided virtual power plant multi-agent collaborative control method, through CC chaos identification + phase space reconstruction + LSTM prediction, can effectively exploit the chaotic characteristics of electricity prices, significantly improve the prediction accuracy for extreme fluctuation scenarios such as price peaks and sharp drops, and avoid market transaction losses caused by price prediction deviations; the average error of medium- and long-term (1-7 days) electricity price prediction is reduced by 21.7%, and the average MAPE is only 9.4%, which is better than traditional LSTM, GRU, and SVR models; In this embodiment of the invention, the provided virtual power plant multi-agent collaborative control method, with its three-level multi-agent collaborative architecture of region-station-equipment, enables hierarchical decision-making and intelligent linkage. The adjustment accuracy error in scenarios such as frequency regulation and peak shaving is controlled within 5%, avoiding performance losses due to adjustment deviations and improving grid access compatibility. The average response time in the frequency regulation scenario is only 8.7 seconds, far below the grid's strict standard of ≤10 seconds, allowing for rapid adaptation to the real-time scheduling needs of the ancillary services market. Through hierarchical optimization and efficient resource integration, the revenue from the frequency regulation scenario increases by 25.0%, the peak shaving scenario by 26.3%, the spot arbitrage scenario by 29.3%, and the average monthly comprehensive revenue increases by 26.9%.
[0019] The virtual power plant multi-agent collaborative control system of the present invention includes the aforementioned virtual power plant multi-agent collaborative control method, and therefore possesses at least all the features and advantages of the aforementioned virtual power plant multi-agent collaborative control method, which will not be repeated here. Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description
[0020] Figure 1 This is a framework diagram of a multi-scale load prediction model disclosed in one embodiment of the present invention; Figure 2 This is a diagram of a multi-agent collaborative control architecture disclosed in one embodiment of the present invention. Detailed Implementation
[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0022] The embodiments of the present invention will be described in detail below with reference to the accompanying drawings and specific examples and application scenarios.
[0023] This invention focuses on two core aspects: accurate prediction and coordinated control. It systematically constructs a multi-agent coordinated control system for virtual power plants based on deep learning. By integrating LSTM and GRU to build a multi-scale load prediction model, it achieves high-precision synchronous prediction of daily, hourly, and minute-level loads. It introduces chaos theory and phase space reconstruction methods to enhance the ability to uncover the inherent patterns in electricity price sequences. Based on this, it designs a three-level multi-agent coordinated architecture of "region-station-equipment" to achieve rapid response and maximized benefits of distributed resources, providing a complete technical solution for the efficient operation of virtual power plants in energy exchange integration scenarios.
[0024] The core is a closed-loop process of data input → dual-model prediction → three-level coordinated regulation → benefit feedback. The positioning and correlation of each research method are as follows: Multi-scale load forecasting module: Improves the accuracy of daily-hourly-minute load forecasting, providing precise load demand input for control decisions; its implementation steps include: large model pre-training → multi-scale model fine-tuning → entropy weight method dynamic fusion, and the output is the load forecasting result; The chaotic-deep learning coupled electricity price prediction module aims to reduce the bias in medium- and long-term electricity price predictions and provide accurate market price inputs for regulatory decisions. Its implementation steps include: chaotic characteristic identification → phase space reconstruction → LSTM prediction, and the output is the electricity price prediction result. The three-level multi-agent collaborative optimization module: Based on the dual prediction results, it realizes the global optimization scheduling of distributed resources (maximizing global benefits and fast response); its implementation steps include: three-level architecture design → objective function construction → constraint setting → closed-loop collaboration; this module depends on the prediction results of the previous two modules and outputs corresponding scheduling instructions.
[0025] In some embodiments of this invention, the prediction layer serves as the information foundation for regulatory decisions, and its accuracy directly affects the subsequent optimization results. This invention constructs a multi-scale load prediction model and a price prediction model based on chaos theory, respectively, to address the different characteristics of load and electricity prices.
[0026] The multi-scale load forecasting model is constructed by fusing a Long Short-Term Memory (LSTM) network with a Gated Recurrent Unit (GRU). Figure 1 As shown. Long-term forecast (daily-weekly): Utilizes an LSTM model, inputting historical load, meteorological, and policy data, outputting the load for the next day to the following week; Short-term forecast (hourly): Utilizes a GRU model, inputting ultra-short-term meteorological data and real-time load data, outputting the load for the next 1-6 hours; Ultra-short-term forecast (15-minute): Utilizes an LSTM-GRU fusion model, determining weights through entropy weighting, outputting the load for the next 15-60 minutes.
[0027] In some embodiments of the present invention, the LSTM cell formula in the multi-scale load forecasting model is as follows: (1) in, , , These are the input gate, forget gate, and output gate, respectively. In cellular state, It is in a hidden state. x t for t The feature vector input at each time step includes historical load. L t-1 Meteorological data W t Holiday signs H t Traffic flow T t wait; h t-1 It is the hidden state of the previous time step, which passes on the temporal dependency information; Ct-1 Record the long-term load trend based on the cell state at the previous moment; W The model weight matrix is obtained through training. b All are bias vectors; output h t Used to predict the load in the next time period L t . This represents the Sigmoid activation function value.
[0028] In some embodiments of the present invention, the GRU cell formula in the multi-scale load forecasting model is as follows: (2) in, This is an update gate used to control the proportion of historical information retained; This is a reset gate used to control the degree to which historical information is forgotten; This is a candidate hidden state, used to combine the current input with historical information; This is the output used to predict the load for the next 1-6 hours.
[0029] In some embodiments of the present invention, the content of the fusion weight calculation (entropy weight method) in the multi-scale load forecasting model is as follows: Define LSTM prediction error GRU prediction error Weights of LSTM prediction results Weights of GRU prediction results , (3) Fusion prediction results: (4) in, P LSTM ( t The result is the LSTM forecast, which is the load forecast value, in megawatts. P GRU ( t The result is the GRU forecast, which is the load forecast value, in megawatts. P pred ( t The result is the combined load forecast, which is used for subsequent control decisions.
[0030] Example 1: Taking the load forecast of a certain CCCC industrial park as an example, the data is as follows: Historical data: Load data from January to December 2024 (15-minute intervals, totaling 35,040 data points); Input characteristics: historical load, irradiance, temperature, holidays, traffic flow (traffic flow on highways surrounding the park); Forecast targets: load for the next 15 minutes, 1 hour, and 24 hours.
[0031] Model parameter settings: LSTM: 2 hidden layers, 128 neurons, learning rate 0.001, 100 iterations; GRU: 2 hidden layers, 64 neurons, learning rate 0.001, 100 iterations.
[0032] The prediction results are shown in Table 1.
[0033] Table 1 Comparison of Multi-Scale Load Forecasting Errors
[0034] Examples of 15-minute predictions are shown in Table 2.
[0035] Table 2 Example of 15-minute load forecast (January 1, 2025)
[0036] The average fusion error was 2.1%, and the load prediction accuracy was 97.9%, which meets the project requirements (≥95%).
[0037] In some embodiments of the present invention, the phase space reconstruction price prediction model, namely the chaos-deep learning coupled electricity price prediction module, includes the following: Electricity spot prices exhibit chaotic characteristics. The CC method is used to determine the phase space reconstruction parameters. (1) Delay time Calculated through the autocorrelation function, , Autocorrelation function; delay time This represents the time interval between adjacent states in the electricity price sequence. (2) Embedding dimension The dimension of the reconstructed phase space is represented by the GP algorithm. This reflects the complexity of the electricity price dynamics system.
[0038] The phase space reconstruction formula is as follows: Let the electricity price time series be... The reconstructed phase space vectors are: (5) in, The number of phase space vectors; x ( t ) is of length M The original electricity price time series, in yuan / MWh ; Reconstructed phase space vectors X ( t ) indicates that the electricity price power system is in m The state evolution trajectory in 3D space, each vector containing m Electricity price data at consecutive time points are used to explore the nonlinear dynamics of electricity prices.
[0039] LSTM-based price prediction model: reconstructing phase space vectors For input, future h Electricity price For the output, an LSTM prediction model is constructed, with the mean squared error (MSE) used as the loss function: (6) Example 2: Taking Shanghai electricity spot price data as an example, the data is as follows: Historical data: Spot prices from July to December 2024 (15-minute timeframe, 17,520 data points in total); Forecast target: Spot price over the next 7 days (15-minute timeframe).
[0040] Parameter calculation: (1) Delay time Autocorrelation function hour, ; (2) Embedding dimension : GP algorithm calculation ; (3) Number of phase space vectors .
[0041] Model parameter settings: LSTM: 3 hidden layers, 256 neurons, learning rate 0.0005, 150 iterations; Predicted step size (1 day = 96 data points at the 15-minute level). The prediction results are shown in Table 3.
[0042] Table 3 Comparison of Electricity Price Forecast Errors
[0043] The average error reduction rate was 21.7%, meeting the project requirements (a 20% reduction). An example of a price forecast for the next day is shown in Table 4.
[0044] Table 4 Example of electricity price forecast for the next day (January 1, 2025)
[0045] In some embodiments of the present invention, the multi-agent cooperative optimization architecture is as follows: The decision-making layer receives information from the forecasting layer and, through multi-agent collaborative optimization, formulates a scheduling plan with the goal of maximizing market returns.
[0046] Multi-Agent Collaborative Control Architecture Design: A three-tiered multi-agent (MAS) architecture is adopted, consisting of a regional-station-equipment level. The regional-level agent embeds a large-scale power time-series model, responsible for global decision-making and market transactions. The station-level agent deploys a deep learning optimizer, responsible for resource scheduling within the region. The equipment-level agent implements real-time control and status feedback. For example... Figure 2 As shown.
[0047] Regional-level intelligent agent (RA): responsible for cross-regional resource coordination, market transaction decision-making, and connecting with the electricity market platform; Site-level intelligent agent (SA): responsible for resource optimization and scheduling within a single site (such as a park or port), receiving RA instructions and reporting the execution status; Device-level intelligent agent (DA): responsible for the real-time control of DER devices (photovoltaic, energy storage, charging piles) and executing SA commands.
[0048] Collaborative optimization objective function: Taking a regional-level intelligent agent as an example, the optimization objective is to maximize market returns, and the constraints include resource constraints, power grid constraints, and policy constraints. (7) in: For demand response revenue; For ancillary service revenue; For arbitrage profits in the spot market; Operating costs.
[0049] (8) in, For the first The adjustment capacity of each resource Demand response electricity price (RMB / MWh) The scheduling cycle consists of 96 15-minute time slots. This represents the total number of resources.
[0050] Taking frequency modulation as an example, (9) in, The unit price for frequency modulation services (RMB / (MW)) h), To adjust the duration.
[0051] (10) In the formula: fort Real-time power consumption (MW); for t Real-time grid connection electricity price (RMB / MWh); for Power purchased at any time (MW); for Real-time electricity purchase price (RMB / MWh).
[0052] The constraints include resource constraints, grid constraints, response time constraints, and energy storage constraints.
[0053] Resource constraints: For the first (Minimum / maximum adjustment capacity of each resource). Grid constraints: (Total grid-connected power in the region ≤ grid limit); Response time constraints: ( For actual response time, frequency modulation scenario ); Energy storage constraints: In the state of energy storage charge, , ).
[0054] In some embodiments of the present invention, the experimental environment and dataset contents are as follows: To verify the effectiveness and practicality of the proposed technical solution, a high-performance computing experimental platform was built. The hardware configuration included an Intel Xeon Gold 6248 CPU, 128GB RAM, and an NVIDIA A100 GPU for accelerating model training. The software environment consisted of Python 3.9, TensorFlow 2.8, and MATLAB 2023b. Experimental data were obtained from the actual operation data of three typical industrial parks under China Communications Construction Group (Shanghai Qingpu, Guangdong Nansha, and Sichuan Chengdu) throughout 2024. This included output / load data for distributed photovoltaic systems, energy storage systems, and electric vehicle charging piles, total park load data, and corresponding meteorological information and electricity spot market price data. Data was collected every 15 minutes.
[0055] In some embodiments of the present invention, the content of prediction model validation and result analysis is as follows: By comparing experiments and scenario tests, we can quantitatively evaluate the performance improvement of each technical module and its contribution to the overall system.
[0056] (1) Accuracy verification of the multi-scale load forecasting model To verify the superiority of the proposed LSTM-GRU fusion model across multiple time scales, the following comparative experiments were conducted: Baseline models: single LSTM model, single GRU model, and traditional time series model (ARIMA); Evaluation metrics: Mean absolute percentage error (MAPE) and root mean square error (RMSE) are used.
[0057] As shown in Table 5, the fusion model proposed in this invention has significantly lower prediction errors than the benchmark model at the 15-minute, 1-hour, and 24-hour scales. In particular, at the ultra-short-term scale, the fusion model effectively combines the ability of LSTM to capture long-term dependencies with the rapid response characteristics of GRU to short-term fluctuations by dynamically adjusting weights through the entropy weight method, improving the accuracy to 97.9% and providing key input for subsequent high-precision regulation.
[0058] Table 5 Comparison of Multi-Scale Load Forecasting Performance
[0059] (2) Validation of the effectiveness of electricity price prediction based on chaos theory To demonstrate the role of phase space reconstruction in revealing the nonlinear chaotic characteristics of electricity prices, the phase space reconstruction + LSTM model is compared with traditional LSTM, GRU, and SVR (Support Vector Regression) models. First, the CC method is used to calculate the delay time τ=6 and the embedding dimension m=8 of the electricity price series, confirming its chaotic characteristics (the maximum Lyapunov exponent is positive). The reconstructed phase space maps the one-dimensional time series to a high-dimensional dynamical system, enabling LSTM to learn richer state evolution patterns.
[0060] In the task of predicting the next 1 to 7 days, the model of this invention achieved an average MAPE of 9.4%, which is 21.7% lower than the traditional LSTM model. It is significantly more accurate in capturing price peaks and sharp drops, as shown in Table 6, demonstrating the core value of chaos theory in improving the ability to characterize electricity price forecasts.
[0061] Table 6 Comparison of Electricity Price Forecast Errors (MAPE, %)
[0062] Comprehensive performance verification of the multi-agent collaborative control architecture: Under the multi-agent collaborative architecture, tests were conducted on three typical scenarios: frequency regulation, peak shaving, and spot arbitrage. The results are shown in Table 7. In the frequency regulation scenario, the average system response time is 8.7 seconds, meeting the stringent requirement of ≤10 seconds for the power grid; the regulation accuracy error in all scenarios is controlled within 5%. In terms of market returns, compared with traditional control methods, the proposed strategy improves the average monthly comprehensive return by 26.9%, fully demonstrating the significant effect of the AI control technology of this invention in enhancing the economic competitiveness of VPPs.
[0063] Table 7. Effects of Multi-Agent Cooperative Regulation
[0064] The overall benefit increased by 26.9%, the response time (8.7s in frequency modulation scenarios) met the project requirements (≤10s), and the adjustment accuracy (≤5%) reached the industry-leading level.
[0065] Any part of this invention not described in detail can be referred to in the prior art or in the art known to those skilled in the art. This embodiment does not limit such part and will not describe it in detail here.
[0066] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of the present invention without departing from the spirit and scope of the claims, and all of these forms are within the protection scope of the present invention.
Claims
1. A method for multi-agent collaborative control of a virtual power plant, characterized in that, The method includes: S1. Perform multi-scale load forecasting and output the forecast results; S2. Perform chaotic-deep learning coupled electricity price prediction and output the prediction results; S3. Optimize resource scheduling based on the prediction results of S1 and S2.
2. The virtual power plant multi-agent collaborative control method according to claim 1, characterized in that, Step S1 includes: S11. Collect data and perform data preprocessing to provide input for the split-scale model; S12. Each scaled model makes its own predictions; S13. The results of each prediction are dynamically fused using entropy weights to obtain the results of multi-scale load prediction.
3. The virtual power plant multi-agent collaborative control method according to claim 2, characterized in that, The scaled model includes an LSTM sub-model, a GRU sub-model, and an entropy weight dynamic fusion model, which respectively fuse long-term prediction, short-term prediction, and long-short-term prediction results.
4. The virtual power plant multi-agent collaborative control method according to claim 3, characterized in that, The entropy weight dynamic fusion includes: calculating the prediction errors of the LSTM sub-model and the GRU sub-model, determining the weights of the two models using the entropy weight method, and obtaining the fused prediction result based on the weights and prediction results as the result of multi-scale load prediction, so as to reduce the error of a single model.
5. The multi-agent collaborative control method for virtual power plants according to claim 2, characterized in that, The data collected in S11 includes multi-dimensional feature data such as historical load data, meteorological data, policy data, holidays, and traffic flow within a specific time period.
6. The virtual power plant multi-agent collaborative control method according to claim 3, characterized in that, The long-term forecast is as follows: an LSTM sub-model is used to capture the long-term load dependence pattern and output the load for the next day to the following week. The short-term forecast is: using the GRU sub-model to quickly respond to short-term load fluctuations and output the load for the next 1-6 hours; The fusion of long-term and short-term forecast results is as follows: based on their respective weights, the results of long-term and short-term forecasts are merged to output the load for the next 15-60 minutes.
7. The virtual power plant multi-agent collaborative control method according to claim 1, characterized in that, Step S2 includes: S21. Chaotic Characteristics Identification: The CC method is used to analyze the electricity price time series, calculate the delay time and embedding dimension, and verify the chaotic characteristics of the electricity price series; S22. Phase Space Reconstruction: Mapping a one-dimensional electricity price time series to a high-dimensional phase space, constructing a phase space vector, and revealing the nonlinear dynamic evolution law of electricity prices; S23.LSTM Model Prediction: Using the reconstructed phase space vector as input, the LSTM model is used for prediction. The prediction uses the mean squared error as the loss function and outputs the 15-minute electricity price for the next 1-7 days.
8. The multi-agent collaborative control method for virtual power plants according to claim 1, characterized in that, Step S3 includes: S31. Architecture Initialization: Build a three-level intelligent agent architecture of region level, site level, and equipment level, and clarify the responsibilities and communication mechanisms of intelligent agents at each level; S32. Objective function and constraint setting: With the maximization of market revenue as the core objective, resource constraints such as upper and lower limits of regulation capacity, grid constraints such as total grid-connected power limit, response time constraints, and energy storage constraints involving the state of charge range are set. The market revenue maximization mentioned above includes demand response revenue, ancillary service revenue, and spot arbitrage revenue, minus operating costs; S33. Closed-loop coordinated scheduling: A regional-level intelligent agent receives load and electricity price forecasts and formulates global market trading strategies and cross-regional resource coordination plans. Station-level intelligent agent: Receives instructions from regional-level intelligent agents, optimizes resource scheduling within a single station, and provides feedback on execution status; Equipment-level intelligent agent: Executes site-level intelligent agent commands to achieve real-time control of DER equipment, including photovoltaic, energy storage, and charging piles, and provides synchronous status feedback; Dynamic iteration: Adjusting scheduling instructions based on real-time feedback information to form a closed loop of prediction-control-feedback-optimization.
9. A virtual power plant multi-agent collaborative control system, characterized in that, The system is capable of implementing the multi-agent collaborative control method for virtual power plants as described in any one of claims 1-8; the system comprises: Multi-scale load forecasting module: It is configured to achieve long-term forecasting through the LSTM submodule, short-term forecasting through the GRU submodule, and dynamically calculate weights through the entropy weight method to fuse the long-term and short-term forecasting results and output the load forecasting results. The chaos-deep learning coupled electricity price prediction module is configured to perform chaotic characteristic identification, phase space reconstruction, and LSTM prediction, thereby outputting multi-period electricity price prediction results; and, The three-level multi-agent collaborative control module is configured to construct objective functions and constraint systems, enabling information interaction, collaborative control, and optimization iteration among agents at the regional, station, and equipment levels.
10. The virtual power plant multi-agent collaborative control system according to claim 9, characterized in that, The chaotic-deep learning coupled electricity price prediction module includes: Chaotic characteristic unit: The delay time and embedding dimension are calculated using the CC method; Phase space reconstruction unit: maps a one-dimensional electricity price sequence into a high-dimensional vector; LSTM prediction model: The reconstructed vector is used as the input of the LSTM model, and the output is the electricity price prediction result for multiple time periods; Error assessment unit: Monitors prediction accuracy in real time using MAPE and RMSE metrics.