Control method of energy storage virtual power plant and control system of energy storage virtual power plant

Abstract

The application discloses a control method and a control system. The control method comprises the following steps: acquiring multi-source data; preprocessing the multi-source data to obtain standardized time series data; weighting and fusing photovoltaic power generation data and wind power generation data in the standardized time series data by using a preset attention weight to obtain fusion feature data; performing feature splicing on the fusion feature data and load power consumption data, time-of-use electricity price data, meteorological data, power grid dispatching instructions and user power consumption behavior data in the standardized time series data to form multivariate input features; inputting the multivariate input features into a preset three-dimensional prediction model, and outputting a power generation power prediction value, a load power prediction value and an electricity price prediction value by the three-dimensional prediction model; generating an energy storage control strategy based on the power generation power prediction value, the load power prediction value and the electricity price prediction value; and controlling the working mode of the energy storage equipment according to the energy storage control strategy.

Description

Technical Field

[0001] This application relates to the field of energy storage virtual power plant technology, specifically to a control method and control system for an energy storage virtual power plant. Background Technology

[0002] The predictive dispatch system of energy storage virtual power plants directly impacts the economics of energy storage device charging and discharging strategies and the grid's regulation capabilities. Current predictive dispatch systems generally employ an architecture where each dimension forecasts independently—photovoltaic power generation, wind power generation, load forecasting, and electricity price forecasting are performed independently by different subsystems. This lack of effective data collaboration and information sharing mechanisms between subsystems often leads to physical and logical conflicts between forecast results from different dimensions. Furthermore, the inherent economic coupling between time-of-use electricity price signals and power generation output and load demand is severed, making it difficult to accurately reflect the actual dispatch needs of energy storage devices. Simultaneously, the impact of meteorological factors on photovoltaic and wind power output is simply treated as a common input feature in current dispatch systems, failing to adaptively enhance features based on the varying importance of meteorological conditions. This results in a significant decrease in prediction accuracy under scenarios of sudden weather changes. Summary of the Invention

[0003] This application provides a control method and control system for an energy storage virtual power plant.

[0004] In a first aspect, this application provides a control method for an energy storage virtual power plant, the energy storage virtual power plant including an energy storage device, the control method comprising: acquiring multi-source data, the multi-source data including photovoltaic power generation data, wind power generation data, load electricity consumption data, time-of-use electricity price data, meteorological data, grid dispatch instructions, and user electricity consumption behavior data; preprocessing the multi-source data to obtain standardized time-series data; weighting and fusing the photovoltaic power generation data and wind power generation data in the standardized time-series data using preset attention weights to obtain fused feature data; concatenating the fused feature data with the load electricity consumption data, time-of-use electricity price data, meteorological data, grid dispatch instructions, and user electricity consumption behavior data in the standardized time-series data to form multivariate input features; inputting the multivariate input features into a preset three-dimensional prediction model, and outputting power generation prediction values, load power prediction values, and electricity price prediction values ​​through the three-dimensional prediction model; generating an energy storage control strategy based on the power generation prediction values, load power prediction values, and electricity price prediction values; and controlling the working mode of the energy storage device according to the energy storage control strategy.

[0005] In some embodiments, acquiring multi-source data includes: acquiring photovoltaic power generation data via a photovoltaic inverter at a first preset acquisition frequency; acquiring wind power generation data via a wind power converter at a second preset acquisition frequency; acquiring load electricity consumption data via smart meters at a third preset acquisition frequency; acquiring time-of-use electricity price data via a power grid data platform at a fourth preset acquisition frequency; acquiring meteorological data via a meteorological station or meteorological data service provider at a fifth preset acquisition frequency; acquiring power grid dispatch instructions via a power grid energy management system; and acquiring user electricity consumption behavior data via smart meters.

[0006] In some embodiments, the preprocessing of the multi-source data includes: uniformly resampling the multi-source data from different acquisition frequencies to the same preset reference frequency to align the multi-source data on the time axis to obtain aligned data; removing outlier data points from the aligned data based on preset criteria to obtain first intermediate data; using linear interpolation to complete data in the first intermediate data with consecutive missing time not exceeding a preset time threshold, and replacing data with historical data from the same period with consecutive missing time exceeding the preset time threshold to obtain second intermediate data; and performing normalization processing on the second intermediate data using a preset standardization method to obtain standardized time-series data.

[0007] In some embodiments, the meteorological factors in the meteorological data include light intensity, wind speed, and temperature; the weighted fusion of photovoltaic power generation data and wind power generation data in the standardized time series data using preset attention weights includes: calculating the preset attention weights using the Softmax function, wherein the preset attention weights include light intensity weight, wind speed weight, and temperature weight, and the sum of the light intensity weight, the wind speed weight, and the temperature weight is 1; and weighting the photovoltaic power generation data and wind power generation data at corresponding times in the standardized time series data using the light intensity weight, the wind speed weight, and the temperature weight respectively to obtain the fused feature data.

[0008] In some embodiments, the three-dimensional prediction model includes a three-layer long short-term memory network. The step of inputting the multivariate input features into a preset three-dimensional prediction model and outputting power generation prediction, load prediction, and electricity price prediction through the three-dimensional prediction model includes: inputting the multivariate input features into the three-layer long short-term memory network, and the three-layer long short-term memory network simultaneously outputting the power generation prediction, the load prediction, and the electricity price prediction.

[0009] In some implementations, each layer of the three-layer long short-term memory network contains 128 hidden units, and the Dropout is set to 0.3; the optimizer of the three-layer long short-term memory network adopts the Adam algorithm, with an initial learning rate of 0.001, a batch size of 32, 500 iterations, and mean squared error as the loss function.

[0010] In some embodiments, the control method further includes: using a preset clustering algorithm to cluster users into a preset number of categories based on their electricity consumption behavior characteristics, thereby obtaining category labels for each user; and embedding the category labels as auxiliary features into the multivariate input features to form updated multivariate input features; the step of inputting the multivariate input features into a preset three-dimensional prediction model and outputting power generation prediction, load prediction, and electricity price prediction values ​​through the three-dimensional prediction model includes: inputting the updated multivariate input features into a preset three-dimensional prediction model and outputting power generation prediction, load prediction, and electricity price prediction values ​​through the three-dimensional prediction model.

[0011] In some embodiments, generating an energy storage control strategy based on the predicted power generation, the predicted load power, and the predicted electricity price includes: generating a charging control strategy when the predicted power generation is greater than the predicted load power and the current electricity price is during a low-price period; generating a discharging control strategy when the predicted power generation is less than the predicted load power and the current electricity price is during a high-price period; and generating a standby or maintenance control strategy when the difference between the predicted power generation and the predicted load power is within a preset range.

[0012] In some embodiments, the control method further includes: determining whether the prediction deviation of any one of the predicted power generation value, the predicted load power value, and the predicted electricity price value exceeds a preset deviation threshold; if it exceeds the threshold, then using the prediction results of the other two prediction dimensions to perform cross-validation and reverse correction on the dimension whose prediction deviation exceeds the preset deviation threshold to obtain a corrected three-dimensional prediction value, and executing the step of generating an energy storage control strategy using the corrected three-dimensional prediction value; if it does not exceed the threshold, then executing the step of generating an energy storage control strategy using the current three-dimensional prediction value.

[0013] In some embodiments, the preset deviation threshold includes: a power generation deviation threshold, a load power deviation threshold, and an electricity price deviation threshold; the step of using the prediction results of the other two prediction dimensions to perform cross-validation and reverse correction on the dimensions whose prediction deviation exceeds the preset deviation threshold to obtain the corrected three-dimensional prediction value includes: when the deviation of the power generation prediction value exceeds the power generation deviation threshold, performing reverse correction using the load power prediction value and the electricity price prediction value; when the deviation of the load power prediction value exceeds the load power deviation threshold, performing reverse correction using the power generation prediction value and the electricity price prediction value; when the deviation of the electricity price prediction value exceeds the electricity price deviation threshold, performing reverse correction using the power generation prediction value and the load power prediction value.

[0014] In some implementations, after the prediction results of the other two prediction dimensions are used to perform cross-validation and reverse correction on the dimension whose prediction deviation exceeds the preset deviation threshold to obtain the corrected three-dimensional prediction value, the control method further includes: dynamically adjusting the learning rate of the three-dimensional prediction model according to the changing trend of the prediction error using a preset reinforcement learning method.

[0015] In some embodiments, the step of dynamically adjusting the learning rate of the three-dimensional prediction model using a preset reinforcement learning method based on the changing trend of the prediction error includes: obtaining the power generation prediction error, load prediction error, and electricity price prediction error for the current prediction period; calculating the sum of the power generation prediction error, the load prediction error, and the electricity price prediction error as the total error; increasing the learning rate of the three-dimensional prediction model by a first preset percentage when the total error shows an increasing trend for a consecutive preset number of prediction periods; and decreasing the learning rate by a second preset percentage when the total error shows a decreasing trend and tends to stabilize for a consecutive preset number of prediction periods; the adjustment range of the learning rate is controlled between 0.001 and 0.01.

[0016] Secondly, this application provides a control system for an energy storage virtual power plant, the control system including a data acquisition module, an edge processor, and an energy storage control module. The data acquisition module is configured to acquire multi-source data. The edge processor is connected to the data acquisition module and configured to execute any of the control methods described above to generate an energy storage control strategy. The energy storage control module is connected to both the edge processor and the energy storage device, and is configured to: receive the energy storage control strategy generated by the edge processor, and control the operating mode of the energy storage device according to the energy storage control strategy.

[0017] The control method and control system of this application uniformly acquire and standardize multi-source data such as photovoltaic power generation data, wind power generation data, load power consumption data, time-of-use electricity price data, meteorological data, power grid dispatch instructions, and user electricity consumption behavior data. It then uses preset attention weights to adaptively weight and fuse meteorologically sensitive photovoltaic and wind power characteristics, and then splices them with standardized time-series data such as load, electricity price, and user behavior into multivariate input features. This achieves three-dimensional collaborative prediction of power generation, load power, and electricity price, fully explores the inherent correlation and economic coupling relationship among the three, and overcomes the technical problems of inconsistent results in various dimensions and the separation of electricity price signals and power generation load in the traditional independent prediction mode.

[0018] Additional aspects and advantages of this application 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 this application. Attached Figure Description

[0019] The above and / or additional aspects and advantages of this application will become apparent and readily understood from the description of the embodiments taken in conjunction with the following drawings, wherein: Figure 1 This is a flowchart illustrating the control method of an energy storage virtual power plant according to some embodiments of this application; Figure 2 This is a schematic diagram of the control system of an energy storage virtual power plant according to some embodiments of this application; Figure 3 This is a flowchart illustrating the process of acquiring multi-source data in the control method of an energy storage virtual power plant according to some embodiments of this application; Figure 4 This is a schematic diagram illustrating the principle of the acquisition module acquiring multi-source data in the control system of an energy storage virtual power plant according to some embodiments of this application; Figure 5 This is a flowchart illustrating the preprocessing of multi-source data in the control method of the energy storage virtual power plant according to some embodiments of this application; Figure 6 This is a schematic diagram of the process of weighted fusion of photovoltaic power generation data and wind power generation data in standardized time series data using preset attention weights in the control method of energy storage virtual power plant of some embodiments of this application. Figure 7 This is a schematic diagram of the structure of a three-dimensional prediction model unit in the control system of an energy storage virtual power plant according to some embodiments of this application; Figure 8 This is a flowchart illustrating the control method of an energy storage virtual power plant according to some embodiments of this application; Figure 9 This is a flowchart illustrating the control method of an energy storage virtual power plant according to some embodiments of this application; Figure 10This is a schematic diagram of the process of generating an energy storage control strategy based on the predicted power generation value, the predicted load power value, and the predicted electricity price value in the control method of the energy storage virtual power plant of some embodiments of this application. Figure 11 This is a flowchart illustrating the control method of an energy storage virtual power plant according to some embodiments of this application; Figure 12 This is a schematic diagram of the structure of the correction and reinforcement learning unit in the control system of an energy storage virtual power plant according to some embodiments of this application; Figure 13 This is a schematic diagram of the process of using the prediction results of the other two prediction dimensions to perform cross-validation and reverse correction on the dimension whose prediction deviation exceeds the preset deviation threshold in the control method of the energy storage virtual power plant of some embodiments of this application. Figure 14 This is a flowchart illustrating the control method of an energy storage virtual power plant according to some embodiments of this application; Figure 15 This is a flowchart illustrating the process of dynamically adjusting the learning rate of a three-dimensional prediction model using a preset reinforcement learning method based on the changing trend of prediction error in the control method of an energy storage virtual power plant according to some embodiments of this application. Figure 16 This is a flowchart illustrating the control method of an energy storage virtual power plant according to some embodiments of this application.

[0020] The reference numerals in the detailed embodiments are as follows: Control system 100; Data Acquisition Module 10, Photovoltaic Data Acquisition Submodule 11, Wind Power Data Acquisition Submodule 12, Load Data Acquisition Submodule 13, Electricity Price Data Acquisition Submodule 14, Meteorological Data Acquisition Submodule 15, Dispatch Instruction Acquisition Submodule 16, User Behavior Acquisition Submodule 17; Edge processor 30, preprocessing module 31, 3D prediction model unit 33, attention weighting subunit 331, feature splicing subunit 333, clustering auxiliary subunit 335, 3D prediction subunit 337, policy generation unit 35, correction and reinforcement learning unit 37, bias detection subunit 371, cross-validation correction subunit 373, learning rate adaptive adjustment subunit 375; Energy storage control module 50; Energy storage device 200, energy storage unit 201, power conversion unit 203. Detailed Implementation

[0021] To make the above-mentioned objectives, features, and advantages of this application more apparent and understandable, the specific embodiments of this application are described in detail below with reference to the accompanying drawings. Many specific details are set forth in the following description to provide a thorough understanding of this application. However, this application can be implemented in many other ways different from those described herein, and those skilled in the art can make similar modifications without departing from the spirit of this application. Therefore, this application is not limited to the specific embodiments disclosed below.

[0022] In the description of this application, it should be understood that the terms "center", "length", "upper", "lower", "front", "rear", "vertical", "horizontal", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this application and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this application.

[0023] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this application, "multiple" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0024] In this application, unless otherwise expressly specified and limited, the terms "installation," "connection," "joining," etc., should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral part; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal communication of two components or the interaction between two components, unless otherwise expressly limited. Those skilled in the art can understand the specific meaning of the above terms in this application according to the specific circumstances.

[0025] In this application, unless otherwise expressly specified and limited, "above" or "below" the second feature can mean that the first feature is in direct contact with the second feature, or that the first feature is in indirect contact with the second feature through an intermediate medium. Furthermore, "above," "on top of," and "over" the second feature can mean that the first feature is directly above or diagonally above the second feature, or simply that the first feature is at a higher horizontal level than the second feature. "Below," "below," and "under" the second feature can mean that the first feature is directly below or diagonally below the second feature, or simply that the first feature is at a lower horizontal level than the second feature.

[0026] The predictive dispatch system of an energy storage virtual power plant directly affects the economics of the energy storage device's charging and discharging strategy and the grid's regulation capability. Current predictive dispatch systems generally adopt an architecture where each dimension of prediction is independent, meaning that photovoltaic (PV) power generation prediction, wind power generation prediction, load prediction, and electricity price prediction are completed independently by different subsystems. The lack of effective data collaboration and information sharing mechanisms between these subsystems often leads to physical and logical conflicts between prediction results from different dimensions. Furthermore, the inherent economic coupling between time-of-use (TOU) electricity price signals and power generation output and load demand is severed, making it difficult to accurately reflect the actual dispatch needs of energy storage devices. Simultaneously, the impact mechanism of meteorological factors on PV and wind power output is simply treated as a common input feature in current dispatch systems, failing to adaptively enhance features based on the varying importance of meteorological conditions. This results in a significant decrease in prediction accuracy under sudden weather changes. To address this issue, this application provides a control method for an energy storage virtual power plant (…). Figure 1 , Figure 3 , Figure 5 , Figure 6 , Figures 8 to 11 and Figures 13 to 16 (as shown) and the control system 100 of the energy storage virtual power plant Figure 2 , Figure 4 , Figure 7 and Figure 12 (As shown).

[0027] Please see Figure 1 and Figure 2 In a first aspect, this application provides a control method for an energy storage virtual power plant, the energy storage virtual power plant including an energy storage device 200, the control method including: 01: Acquire multi-source data, including photovoltaic power generation data, wind power generation data, load power consumption data, time-of-use electricity price data, meteorological data, power grid dispatch instructions, and user electricity consumption behavior data; 02: Preprocess the multi-source data to obtain standardized time-series data; 03: Use preset attention weights to perform weighted fusion of photovoltaic power generation data and wind power generation data in standardized time series data to obtain fused feature data; 04: The fusion feature data is combined with load electricity consumption data, time-of-use electricity price data, meteorological data, power grid dispatch instructions and user electricity consumption behavior data from the standardized time series data to form multivariate input features; 07: Input multivariate input features into a preset three-dimensional prediction model, and output the predicted power generation, predicted load, and predicted electricity price through the three-dimensional prediction model; 09: Generate energy storage control strategies based on predicted power generation, predicted load, and predicted electricity prices; and, 010: Control the working mode of energy storage device 200 according to the energy storage control strategy.

[0028] Correspondingly, in a second aspect, this application also provides a control system 100 for an energy storage virtual power plant. The control system 100 includes a data acquisition module 10, an edge processor 30, and an energy storage control module 50. The data acquisition module 10 is configured to acquire multi-source data. The edge processor 30 is connected to the data acquisition module 10 and is configured to execute control methods 01, 02, 03, 04, 07, and 09 to generate an energy storage control strategy. The energy storage control module 50 is connected to the edge processor 30 and the energy storage device 200, respectively, and is configured to receive the energy storage control strategy generated by the edge processor 30, and control the operating mode of the energy storage device 200 according to the energy storage control strategy; that is, the energy storage control module 50 is used to execute the control method 010.

[0029] A virtual energy storage power plant is a structure that aggregates distributed energy storage resources into a unified whole through a communication network and a control system 100, participating in grid operation scheduling and electricity market transactions. The virtual energy storage power plant itself does not rely on centralized large-scale energy storage power plants, but rather logically integrates multiple dispersed energy storage devices 200 into a unified dispatchable resource, subject to unified management and allocation by the control system 100. The control system 100 is the decision-making center and execution hub of the virtual energy storage power plant, responsible for the complete closed-loop process from data acquisition, data processing, predictive calculation to control command issuance. The energy storage devices 200 are the direct targets of the control system 100, and their operating modes include at least charging mode, discharging mode, and standby mode. The switching between these three modes and the adjustment of power output are determined by commands issued by the energy storage control module 50.

[0030] The data acquisition module 10 serves as the data entry point for the control system 100, responsible for acquiring various raw data required for the normal operation of the control system 100 from external data sources. Physically, the acquisition module 10 can be a communication interface board integrated within the edge processor 30, or it can be a data acquisition terminal independent of the edge processor 30. The acquisition module 10 connects to the edge processor 30 via industrial Ethernet or an internal bus, transmitting the acquired multi-source data to the edge processor 30. The acquisition module 10 must support multiple physical interfaces and communication protocols to adapt to the differences in interface form, electrical characteristics, and communication protocols among different types of data sources.

[0031] Photovoltaic power generation data is time-series data characterizing the real-time power generation of a photovoltaic power plant. The photovoltaic power generation data originates from the photovoltaic inverter, which, while performing DC-AC conversion, integrates power measurement and communication modules, enabling it to output the measured value of the current power generation in real time. The acquisition module 10 establishes a communication connection with the photovoltaic inverter via an RS485 bus or Ethernet, and reads the power data according to a preset acquisition frequency.

[0032] Wind power generation data is time-series data characterizing the real-time power generation of a wind farm. This data originates from the wind power converter, which, while performing frequency conversion and grid connection control, also incorporates a power measurement unit and a communication protocol stack. The acquisition module 10 communicates with the wind power converter via the IEC 61850 standard protocol, reading power data according to a preset acquisition frequency.

[0033] Load power consumption data is time-series data characterizing the real-time power consumption of users within the coverage area of ​​the energy storage virtual power plant. The load power consumption data originates from smart meters, which are installed at the incoming terminals of each user or the outgoing terminals of each feeder, continuously measuring voltage and current and calculating active power. The acquisition module 10 communicates with the smart meters via the DL / T645 protocol and reads the load power data according to a preset acquisition frequency.

[0034] Time-of-use (TOU) electricity price data refers to the electricity price information for different time periods released by the power grid company. This TOU data originates from the power grid data platform, which is operated by the power grid company and provides standardized data service interfaces. The data acquisition module 10 sends requests to the power grid data platform via API calls to obtain the electricity price time periods and corresponding price values ​​for the current day or several future days.

[0035] Meteorological data characterizes the environmental conditions of the area where the energy storage virtual power plant is located. Meteorological data originates from meteorological stations deployed near photovoltaic power plants and wind farms, or from third-party meteorological data service providers. The data includes parameters such as sunlight intensity, wind speed, and temperature. These parameters directly affect the photoelectric conversion efficiency of photovoltaic modules and the operating characteristics of wind turbines, and are key input variables for predicting photovoltaic and wind power generation. The acquisition module 10 connects to the meteorological station via serial or network communication, or calls the meteorological service provider's API interface via HTTP protocol, acquiring meteorological data at a preset acquisition frequency.

[0036] Power grid dispatch instructions are adjustment commands issued by the power grid dispatch center based on the power grid's operating status. These instructions originate from the power grid energy management system, which is responsible for real-time monitoring and dispatch decisions. The acquisition module 10 communicates with the power grid energy management system via the MMS protocol to acquire power adjustment instructions, operating mode instructions, and emergency control instructions in real time.

[0037] User electricity consumption behavior data is data that characterizes users' electricity consumption habits and patterns. This data originates from the historical records of smart meters. While continuously measuring load power, smart meters store the measurement data internally or upload it to a data collection system. The acquisition module 10 obtains characteristic data such as users' daily electricity consumption curves, peak-valley electricity consumption ratios, and weekly electricity consumption pattern differences from the smart meters or the data collection system.

[0038] Edge processor 30 is the computing core of control system 100, deployed on the field side of the energy storage virtual power plant. Edge processor 30 connects to acquisition module 10, receives multi-source data transmitted from acquisition module 10, and performs computational tasks locally, including data preprocessing, feature fusion, 3D prediction, and control strategy generation. The difference between edge processor 30 and a cloud server is that edge processor 30 is deployed physically close to the data source and device, allowing data to be processed locally without wide area network transmission, thus shortening the time delay between data acquisition and control command issuance. Edge processor 30 can be implemented using an industrial-grade edge computing server or an embedded industrial control computer, integrating a processor, memory, storage, and communication interface.

[0039] The energy storage control module 50 is the execution layer of the control system 100, and is connected to both the edge processor 30 and the energy storage device 200. The energy storage control module 50 receives the energy storage control strategy generated by the edge processor 30 and converts it into drive signals for the energy storage device 200. The energy storage control module 50 and the edge processor 30 are connected via fiber optic Ethernet to ensure low-latency transmission of control commands. The energy storage control module 50 and the energy storage device 200 are connected via hardwiring or fieldbus, outputting PWM pulse signals or analog commands to directly control the power conversion unit 203 of the energy storage device 200.

[0040] In step 01: the acquisition module 10 acquires photovoltaic power generation data, wind power generation data, load power consumption data, time-of-use electricity price data, meteorological data, grid dispatch instructions and user electricity consumption behavior data according to their respective adapted acquisition frequencies. These seven types of data reflect the operating environment and status of the energy storage virtual power plant from five dimensions: generation side, load side, market side, environment side and dispatch side.

[0041] In step 02: Preprocessing involves cleaning and formatting the acquired multi-source data. Since the seven types of data come from different data sources, their collection frequencies, units, data completeness, and time bases vary. Directly retaining these differences would lead to biases in subsequent feature fusion and model prediction. Preprocessing unifies data from different frequencies to the same time base, removes erroneous data that clearly exceeds reasonable limits, completes missing data, and performs standardization on all data to eliminate the influence of units, outputting standardized time-series data with a unified time axis, unified units, and complete data points.

[0042] In step 03: Attention weights are quantitative coefficients used to measure the influence of different meteorological factors on photovoltaic (PV) and wind power generation. In PV and wind power generation forecasting, the influence of the three meteorological factors—solar intensity, wind speed, and temperature—varies. Solar intensity is more important than wind speed in PV forecasting, wind speed is more important than solar intensity in wind power forecasting, and temperature has a relatively minor impact on both types of power generation. The preset values ​​for attention weights are calculated using the Softmax function, ensuring that each weight value is between 0 and 1 and that the sum is 1 (described below). Using the preset attention weights, the PV and wind power generation data in the standardized time-series data are weighted and fused. This involves multiplying the weights of each meteorological factor by the corresponding PV and wind power generation characteristics at that time and then summing the results to obtain a fused power generation feature vector. In this vector, feature components strongly correlated with current meteorological conditions are amplified, while weakly correlated feature components are suppressed, thus achieving adaptive enhancement of meteorologically sensitive features.

[0043] In step 04: Feature concatenation is the operation of merging multiple data vectors into a single higher-dimensional vector along the feature dimensions. The fused feature data represents the generation-side features after meteorological attention weighting. The load consumption data, time-of-use electricity price data, meteorological data, grid dispatch instructions, and user electricity behavior data in the standardized time-series data represent load-side features, market-side features, environmental-side features, dispatch-side features, and behavioral-side features, respectively. Through feature concatenation, the generation-side features are merged with the features from the other five aspects into multivariate input features, enabling the subsequent prediction model to simultaneously obtain complete information from six dimensions.

[0044] In step 07: The three-dimensional prediction model is a pre-trained deep learning model. Its input consists of multivariate input features, and its output consists of predicted values ​​in three dimensions: power generation, load, and electricity price. These three dimensions are not independent of each other; there is an economic coupling relationship between electricity price and load demand, and a supply-demand balance relationship between load demand and renewable energy generation. These inherent connections are learned and solidified by the model parameters during the training process. After the three-dimensional prediction model outputs the three predicted values, these values ​​can serve as the decision-making basis for generating energy storage control strategies.

[0045] In step 09: The energy storage control strategy is a decision-making scheme generated based on the predicted power generation, predicted load power, and predicted electricity price. Specifically, the relationship between the predicted power generation and predicted load power is compared to determine the direction of the power difference of the virtual power plant in the future period. At the same time, the electricity price prediction is referenced to determine whether the current period is a low-price or high-price period. The direction of the power difference and the electricity price signal are combined to determine whether the energy storage device 200 should perform a charging operation or a discharging operation.

[0046] The energy storage control module 50 controls the operating mode of the energy storage device 200 according to the energy storage control strategy generated by the edge processor 30. Specifically, the energy storage control module 50 analyzes the target power value and operating mode parameters in the energy storage control strategy, generates a corresponding drive signal through its internal control algorithm, and outputs the drive signal to the power conversion unit 203 of the energy storage device 200. The power conversion unit 203 adjusts the charging and discharging current and voltage of the energy storage unit 201 according to the drive signal, so that the energy storage device 200 outputs or absorbs power according to the strategy requirements.

[0047] The control method and control system 100 of this application uniformly acquire and standardize multi-source data such as photovoltaic power generation data, wind power generation data, load power consumption data, time-of-use electricity price data, meteorological data, power grid dispatch instructions, and user electricity consumption behavior data. It then uses preset attention weights to adaptively weight and fuse meteorologically sensitive photovoltaic and wind power characteristics, and then splices them with standardized time-series data such as load, electricity price, and user behavior into multivariate input features. This achieves three-dimensional collaborative prediction of power generation, load power, and electricity price, fully explores the inherent correlation and economic coupling relationship among the three, and overcomes the technical problems of inconsistent results in various dimensions and the separation of electricity price signals and power generation load in the traditional independent prediction mode.

[0048] Please see Figure 3 and Figure 4 In some implementations, 01: Acquiring multi-source data includes: 011: Photovoltaic power generation data is collected via a photovoltaic inverter at a first preset acquisition frequency; 012: Collect wind power generation data through the wind power converter at the second preset acquisition frequency; 013: Collect load power consumption data through smart meters according to the third preset collection frequency; 014: Time-of-use electricity price data is acquired through the power grid data platform at the fourth preset acquisition frequency; 015: Meteorological data is acquired through weather stations or meteorological data service providers according to the fifth preset collection frequency; 016: Obtain grid dispatch instructions through the grid energy management system; 017: Collect user electricity consumption data through smart meters.

[0049] Correspondingly, the data acquisition module 10 includes a photovoltaic data acquisition submodule 11, a wind power data acquisition submodule 12, a load data acquisition submodule 13, an electricity price data acquisition submodule 14, a meteorological data acquisition submodule 15, a dispatch instruction acquisition submodule 16, and a user behavior acquisition submodule 17. The photovoltaic data acquisition submodule 11 is configured to acquire photovoltaic power generation data via a photovoltaic inverter at a first preset acquisition frequency. The wind power data acquisition submodule 12 is configured to acquire wind power generation data via a wind power converter at a second preset acquisition frequency. The load data acquisition submodule 13 is configured to acquire load electricity consumption data via smart meters at a third preset acquisition frequency. The electricity price data acquisition submodule 14 is configured to acquire time-of-use electricity price data via a power grid data platform at a fourth preset acquisition frequency. The meteorological data acquisition submodule 15 is configured to acquire meteorological data via a meteorological station or meteorological data service provider at a fifth preset acquisition frequency. The dispatch instruction acquisition submodule 16 is configured to acquire power grid dispatch instructions via a power grid energy management system. The user behavior acquisition submodule 17 is configured to acquire user electricity consumption behavior data via smart meters.

[0050] A photovoltaic (PV) inverter is the core power electronic device in a photovoltaic (PV) power generation system, responsible for converting the direct current (DC) generated by PV modules into alternating current (AC) and feeding it into the power grid. While performing power conversion, the PV inverter integrates voltage and current sampling circuits, an analog-to-digital converter (ADC), and a digital signal processor (DSP). It can measure the output voltage, output current, and output power of the PV array in real time and store these measurements in internal registers for external retrieval. The PV inverter provides an RS485 or Ethernet communication interface, supporting data transmission via Modbus or a custom protocol. The PV data acquisition submodule 11 establishes a physical connection with the PV inverter via RS485 bus or Ethernet, sending data read requests to the PV inverter at a first preset acquisition frequency. Upon receiving the request, the PV inverter replies with its current power data in the form of data frames. The PV data acquisition submodule 11 then parses the PV power generation value from the replied data frames. In one example, the first preset acquisition frequency is set to once per minute. This frequency is adapted to the rate of change of photovoltaic power generation. Photovoltaic power generation is affected by cloud movement and changes in solar irradiance. Its time scale of change is usually from several minutes to tens of minutes. Acquiring once per minute can capture the fluctuation trend of power generation, while not generating too much data storage and transmission burden.

[0051] The wind power converter is the core power electronic device in a wind power generation system, responsible for converting the frequency-converted AC power output from the wind turbine into AC power with the same frequency and phase as the power grid. While performing power conversion, the wind power converter integrates voltage and current sampling circuits and a digital signal processor, enabling real-time measurement of the wind turbine's output power and storing the measured values ​​in internal registers for external retrieval. The wind power converter provides an IEC 61850 standard communication interface, supporting data transmission via the IEC 61850-8-1 MMS protocol. The wind power data acquisition submodule 12 establishes a communication connection with the wind power converter via the IEC 61850 standard protocol, reading power data from the wind power converter's data model at a second preset acquisition frequency. In one example, the second preset acquisition frequency is set to once per minute. This frequency is adapted to the rate of change in wind power generation. Wind power generation is affected by wind speed changes, which typically occur over time scales ranging from several minutes to tens of minutes. Acquiring data once per minute captures the fluctuation trend of power generation.

[0052] Smart meters are electricity metering devices that measure and record users' electricity consumption. Internally, smart meters integrate voltage and current sampling circuits and metering chips, enabling real-time measurement of electrical parameters such as voltage, current, active power, reactive power, and power factor. These parameters are stored in internal registers for external retrieval. Smart meters provide an RS485 communication interface, supporting the DL / T645 communication protocol. The load data acquisition submodule 13 establishes a physical connection with the smart meter via the RS485 bus, sending data read requests to the smart meter at a third preset acquisition frequency. Upon receiving the request, the smart meter replies with the current load power data in the form of data frames. The load data acquisition submodule 13 parses the load power value from the replied data frames. In one example, the third preset acquisition frequency is set to once every 15 minutes. The rate of change in load power is relatively slow, mainly influenced by user work and rest schedules, production plans, and seasonal factors. Acquiring data every 15 minutes can fully reflect the daily load variation curve while avoiding the pressure on the communication bus and data storage caused by excessively high acquisition frequencies.

[0053] The power grid data platform is a data service platform built and operated by the power grid company, providing standardized data service interfaces to the outside world. The power grid data platform stores and manages various types of data related to power grid operation, including time-of-use electricity price information, power grid load information, and power generation plan information. The electricity price data acquisition submodule 14 establishes a connection with the power grid data platform via the network and sends HTTP requests to the power grid data platform's interface via API calls according to the fourth preset collection frequency. The requests carry authentication information and data query parameters. After receiving the request, the power grid data platform verifies the identity and queries the database, returning the retrieved time-of-use electricity price data in JSON format encapsulated in the HTTP response. In one example, the fourth preset collection frequency is set to once per hour. Time-of-use electricity price data is usually updated once a day at midnight, and its time period division and corresponding electricity price remain unchanged within the day. Hourly acquisition ensures the real-time nature of the electricity price data in the local cache while avoiding excessive query pressure on the power grid data platform.

[0054] A weather station is a meteorological monitoring device deployed near photovoltaic power plants and wind farms, responsible for measuring and recording on-site environmental meteorological parameters. The weather station integrates sensors for light intensity, wind speed, temperature, humidity, and air pressure, enabling real-time measurement of meteorological parameters such as light intensity, wind speed, temperature, humidity, and air pressure, and outputting these parameters through a data interface. The weather station provides a serial communication interface or a network communication interface, supporting data transmission via Modbus or a custom protocol. The meteorological data acquisition submodule 15 establishes a connection with the weather station via serial or network communication and reads meteorological data according to the fifth preset acquisition frequency. In scenarios where an on-site weather station cannot be deployed, the meteorological data acquisition submodule 15 can obtain meteorological data by calling the API interface of a meteorological data service provider via the HTTP protocol. In one example, the fifth preset acquisition frequency is set to once every 30 minutes. The rate of change of meteorological parameters is relatively slow; light intensity changes gradually with the solar altitude angle, and the timescale of wind speed change due to atmospheric circulation is typically from tens of minutes to several hours. Acquiring data once every 30 minutes can completely record the changing trends of meteorological parameters while avoiding increased communication power consumption caused by excessively high acquisition frequencies.

[0055] The power grid energy management system is the core control system of the power grid dispatch center, responsible for real-time monitoring of the power grid's operating status, dispatch decision-making, and command issuance. The power grid energy management system is connected to each power plant, substation, and energy storage power station via a dedicated dispatch data network, exchanging data in real-time using the MMS protocol. The dispatch command acquisition submodule 16 establishes a communication connection with the power grid energy management system via the MMS protocol, receiving power adjustment commands, operating mode commands, and emergency control commands issued by the power grid energy management system in real time.

[0056] The collection of user electricity consumption data is also accomplished through smart meters. While continuously measuring and recording load electricity data, the smart meter stores historical load data in its internal memory, forming a daily electricity consumption curve record for the user. The user behavior acquisition submodule 17 establishes a physical connection with the smart meter via an RS485 bus, reads the historical load data records stored in the smart meter, and extracts characteristic data such as the user's daily electricity consumption curve shape, peak-valley electricity consumption ratio, weekly electricity consumption pattern differences, and seasonal variation patterns.

[0057] In a specific implementation plan, the first and second preset acquisition frequencies are both once per minute. The interface protocol for photovoltaic power generation data adopts RS485 or Ethernet, while the interface protocol for wind power generation data adopts the IEC61850 standard. The third preset acquisition frequency is once every 15 minutes, with the interface protocol adopting the DL / T 645 standard. The fourth preset acquisition frequency is once per hour, obtained via API calls. The fifth preset acquisition frequency is once every 30 minutes. The interface protocol for grid dispatch instructions adopts the MMS protocol. Each data source acquires raw data of different dimensions according to its adapted acquisition frequency and communication protocol, fully considering the differences in physical characteristics and transmission methods of seven types of data sources: photovoltaic power generation, wind power generation, load power consumption, time-of-use pricing, meteorology, dispatch instructions, and user electricity consumption behavior. Photovoltaic and wind power generation power changes rapidly, requiring a higher acquisition frequency to capture fluctuation details; load power consumption changes relatively smoothly, and a moderate acquisition frequency is sufficient; electricity price data has the lowest update frequency, and an excessively high acquisition frequency does not improve prediction accuracy but wastes communication resources; meteorological data changes gradually, and a moderate acquisition frequency is sufficient to record the complete trend. By configuring differentiated acquisition frequencies and adapted communication protocols for each type of data source, optimal configuration of communication bandwidth and storage resources was achieved while ensuring data integrity and timeliness, providing a comprehensive and reliable data foundation for subsequent 3D collaborative prediction.

[0058] Please see Figure 2 and Figure 5 In some implementations, 02: preprocessing the multi-source data includes: 021: Resample multi-source data from different acquisition frequencies to the same preset reference frequency to align the multi-source data on the time axis and obtain aligned data; 023: Remove outlier data points from the aligned data based on preset criteria to obtain the first intermediate data; 025: For data in the first intermediate data where the consecutive missing time does not exceed a preset time threshold, linear interpolation is used to complete the data; for data where the consecutive missing time exceeds the preset time threshold, historical data from the same period is used to replace the missing data to obtain the second intermediate data; and, 027: The second intermediate data is normalized using a preset standardization method to obtain standardized time series data.

[0059] Correspondingly, the edge processor 30 includes a preprocessing module 31, which is configured to perform steps 021, 023, 025 and 027.

[0060] The preprocessing module 31 is a functional unit within the edge processor 30, responsible for converting the multi-source raw data acquired by the acquisition module 10 into standardized time-series data suitable for the input of the 3D prediction model. The preprocessing module 31 is implemented within the edge processor 30 as a software program module or a hardware acceleration unit. After receiving multi-source data from the acquisition module 10, it sequentially performs resampling, anomaly removal, missing data completion, and standardization operations according to a fixed processing pipeline.

[0061] In step 021, resampling is the operation of unifying multi-source data with different acquisition frequencies to the same reference frequency. The seven types of data acquired by the acquisition module 10 each have different original acquisition frequencies: photovoltaic power generation data and wind power generation data are acquired once per minute, load electricity consumption data is acquired once every 15 minutes, time-of-use electricity price data is acquired once per hour, meteorological data is acquired once every 30 minutes, grid dispatch instructions are event-triggered, and user electricity consumption behavior data is updated daily. The sampling points of these data are unevenly distributed on the time axis, making it impossible to directly perform vectorized feature stitching. Resampling uses a unified time grid as a reference, and calculates the values ​​of the original data at each time grid point through interpolation or forward padding, so that all data have a corresponding value at each time grid point. In one example, the preset reference frequency is once per minute, which is consistent with the highest frequency among all original acquisition frequencies, and can retain the most subtle fluctuation information in photovoltaic power generation data and wind power generation data, while not generating excessively high data density. The resampling process is implemented as follows: a continuous time axis is constructed with each minute as a time node. For a given time node, if a certain type of data has an original sampled value at that node, that original sampled value is used directly; if a certain type of data does not have an original sampled value at that node, the most recent original sampled value before that node is taken as the data value for that node, i.e., forward padding. This method is suitable for data with low collection frequency but whose values ​​remain stable for a short period of time, such as load electricity data and time-of-use electricity price data. For parameters with continuously changing characteristics, such as light intensity in meteorological data, a linear interpolation method is used, i.e., the original sampled points before and after the time node are taken, and the value of that node is calculated by linear weighting according to the time distance. After resampling, all seven types of data have corresponding values ​​at each minute time node, and the multi-source data achieves strict alignment on the time axis, resulting in aligned data output.

[0062] In step 023, outlier removal is an operation that identifies and deletes data points in the aligned data that clearly do not conform to the normal range of variation based on preset criteria. During the data acquisition process, the acquisition module 10 may acquire erroneous data points that are far from the normal value range due to sensor failure, communication interference, signal loss, or instantaneous disturbances. If these outlier data points are retained in the dataset, they will pollute subsequent feature statistics and affect the prediction accuracy of the model. In one example, the preset criterion can be the 3σ criterion, that is, for a set of data that follows a normal or approximately normal distribution, the probability of a data point falling outside the mean ± 3 times the standard deviation is about 0.27%, which is a low-probability event. Therefore, data points that exceed this range can be identified as outliers. The preprocessing module 31 calculates the mean and standard deviation for each type of data, compares each data point with the mean of the data type, and if the absolute value of the difference between the data point and the mean is greater than 3 times the standard deviation, the data point is identified as an outlier and removed from the dataset. After removing outlier data points, the data at that position becomes missing and is handled by the subsequent missing data completion step. The first intermediate data is obtained after anomaly removal.

[0063] In step 025, missing data completion involves filling in the missing values ​​in the first intermediate data caused by anomaly removal. In one example, the preset time threshold is set to 2 hours. Since load and power generation typically do not change abruptly within 2 hours, their changes exhibit good temporal continuity, and interpolation using data from before and after these times can reasonably infer the missing values. For data with consecutive missing periods not exceeding 2 hours, linear interpolation is used for completion, i.e., taking two valid data points before and after the missing segment and linearly calculating the values ​​of each time node within the missing segment according to the time distance. For data with consecutive missing periods exceeding 2 hours, historical data from the same period is used as a replacement, i.e., taking the data value from the same time of the previous day as the replacement value, or taking the average of the data values ​​from the same time of the previous week as the replacement value. After missing data completion, the second intermediate data is obtained.

[0064] In step 027, data standardization is an operation that normalizes the second intermediate data to eliminate dimensional differences. The original dimensions of the seven data categories are different: photovoltaic power generation and wind power generation data are in kilowatts or megawatts; load electricity consumption data is in kilowatts; time-of-use electricity price data is in yuan / kWh; meteorological data includes light intensity in watts / square meter, wind speed in meters / second, and temperature in degrees Celsius; grid dispatch instructions in kilowatts; and user electricity consumption behavior data varies depending on the feature type. Data with different dimensions differ significantly in numerical range. If these data are directly input into a 3D prediction model, features with large numerical ranges will dominate the loss function calculation during model training, making it difficult for features with small numerical ranges to be effectively learned by the model. In one example, the standardization method is Z-score standardization, specifically: for each data point, subtract the mean of that data category and divide by the standard deviation of that data category. After transformation, the mean of each data category is 0, the standard deviation is 1, and the data distribution is converted to a standard normal distribution. Z-score standardization preserves the distribution shape and relative magnitude of the original data while eliminating the influence of units and numerical ranges. After Z-score standardization, all data participate in subsequent feature fusion and model calculations on a uniform numerical scale, outputting standardized time-series data.

[0065] This application aligns multi-source data on the time axis through resampling, and then performs anomaly removal, missing data completion, and standardization in sequence, eliminating differences in collection frequency, units, and completeness among various data sources, thus ensuring the quality and consistency of the data input to the 3D prediction model.

[0066] Please see Figure 6 and Figure 7 In some implementations, the meteorological factors in the meteorological data include light intensity, wind speed, and air temperature; step 03 specifically includes: 031: The preset attention weights are calculated using the Softmax function. These preset attention weights include light intensity weights, wind speed weights, and temperature weights, and the sum of these weights is 1; and... 033: The photovoltaic power generation data and wind power generation data at corresponding times in the standardized time series data are weighted by light intensity weight, wind speed weight and temperature weight respectively to obtain fused feature data.

[0067] Correspondingly, the edge processor 30 includes a three-dimensional prediction model unit 33, which includes an attention weighting subunit 331. The attention weighting subunit 331 is configured to calculate the attention weights of meteorological factors using a Softmax function; and to use these attention weights to perform weighted fusion of photovoltaic power generation data and wind power generation data in the standardized time-series data to obtain fused feature data. The meteorological factors include light intensity, wind speed, and temperature, and the attention weights include light intensity weight, wind speed weight, and temperature weight, the sum of which is 1.

[0068] The Softmax function is a mathematical transformation function used to convert a vector of arbitrary real values ​​into a probability distribution vector, where each element of the transformed vector is between 0 and 1, and the sum of all elements is 1. In step 031, the input to the Softmax function is the original or normalized values ​​of three meteorological factors: light intensity, wind speed, and temperature. The output is three weight values, corresponding to the weights for light intensity, wind speed, and temperature, respectively, and the sum of the three weight values ​​is 1.

[0069] Attention weights are quantitative coefficients used to measure the influence of various meteorological factors on photovoltaic (PV) and wind power generation. Sunlight intensity weight measures the combined influence of sunlight intensity on PV and wind power generation; wind speed weight measures the combined influence of wind speed; and temperature weight measures the combined influence of temperature. When a meteorological factor has a large weight, it indicates that the factor has a dominant influence on power generation at the current moment, and the corresponding power generation characteristic should receive a larger weight in the fused characteristics. Conversely, when a meteorological factor has a small weight, it indicates that the factor has a small influence on power generation at the current moment, and the corresponding power generation characteristic only needs to retain a small weight in the fused characteristics. Sunlight intensity weight primarily affects the weighting of PV power generation data, with little impact on the weighting of wind power generation data. Wind speed weight primarily affects the weighting of wind power generation data, with little impact on the weighting of PV power generation data. Temperature has a secondary influence on both power generation methods, with a lower impact than sunlight intensity on PV and wind speed on wind power.

[0070] In step 033, the photovoltaic (PV) and wind power generation data at the corresponding time point in the standardized time series data are weighted using irradiance weight, wind speed weight, and temperature weight. Specifically, for each time point, the irradiance weight, wind speed weight, and temperature weight values ​​are obtained, as are the PV power generation and wind power generation values. The irradiance weight is multiplied by the PV power generation value, and the wind speed weight and temperature weight are multiplied by the wind power generation value respectively. The three products are then added together to obtain the fused power generation characteristic value for that time point. In some embodiments, the temperature weight is applied to both PV and wind power generation data simultaneously; that is, the temperature weight is multiplied by both the PV power generation value and the wind power generation value and then included in the fused characteristic.

[0071] Through the aforementioned weighting operations, the feature components strongly correlated with current meteorological conditions in the fused feature data are amplified, while weakly correlated feature components are suppressed. Specifically, when the light intensity is high, the weight of light intensity increases accordingly, increasing the proportion of photovoltaic power generation features in the fused features. The model then pays more attention to the historical trends and current values ​​of photovoltaic power generation when predicting power output. When the wind speed is high, the weight of wind speed increases accordingly, increasing the proportion of wind power generation features in the fused features. The model then pays more attention to the historical trends and current values ​​of wind power generation when predicting power output. When the temperature deviates from the suitable range, the weight of temperature increases accordingly, increasing the proportion of the temperature correction term in the fused features. The model then incorporates the impact of temperature on power generation efficiency when predicting power output.

[0072] The specific values ​​of the light intensity weight, wind speed weight, and temperature weight are dynamically calculated based on the meteorological data at each moment. Therefore, the attention weight is a dynamic weight that changes over time, rather than a fixed constant. At different times, due to the different measured values ​​of light intensity, wind speed, and temperature, the three weight values ​​output by the Softmax function also differ. At noon when light intensity is high, the light intensity weight approaches its maximum value, while the wind speed and temperature weights decrease accordingly. At night when light intensity is zero, the light intensity weight approaches zero, and the wind speed and temperature weights are redistributed. During sudden gusts of wind, the wind speed weight increases instantaneously, while the light intensity and temperature weights decrease accordingly. This dynamic weighting mechanism enables the attention-weighted subunit 331 to adaptively adjust the influence distribution of each meteorological factor according to real-time changes in meteorological conditions. This allows the three-dimensional prediction model to focus on the most important meteorological influencing factors under different meteorological scenarios, thereby improving the accuracy of new energy power generation prediction under sudden meteorological changes.

[0073] Please see Figure 1 and Figure 2In one embodiment, the three-dimensional prediction model unit 33 further includes a feature stitching subunit 333. The feature stitching subunit 333 is configured to stitch the fused feature data output by the attention-weighted subunit 331 with load power consumption data, time-of-use electricity price data, meteorological data, power grid dispatch instructions, and user power consumption behavior data in the standardized time series data to form multivariate input features.

[0074] After the attention-weighted subunit 331 completes the weighted fusion of photovoltaic power generation data and wind power generation data, it outputs fused feature data. The fused feature data represents the characteristic information of the power generation side, but the power generation side characteristics are only one aspect of the operating status of the energy storage virtual power plant. Power generation side characteristics alone cannot fully reflect the system's supply and demand balance and the constraints of dispatch decisions. The feature splicing subunit 333 merges the fused feature data with the standardized time-series data of the other five dimensions to construct a comprehensive feature vector containing complete operational information.

[0075] Feature concatenation is an operation that merges multiple independent data vectors into a single vector of higher dimensionality along the feature dimension. This operation does not change the internal values ​​of each vector; it only concatenates the vectors end-to-end along the feature dimension. Mathematically, suppose the fused feature data is an m-dimensional vector, the load electricity consumption data is an n1-dimensional vector, the time-of-use electricity price data is an n2-dimensional vector, the meteorological data is an n3-dimensional vector, the power grid dispatch instructions are an n4-dimensional vector, and the user electricity behavior data is an n5-dimensional vector. After feature concatenation, the multivariate input features become a high-dimensional vector with dimensions m+n1+n2+n3+n4+n5. Feature concatenation requires that the data vectors involved in the concatenation be strictly aligned on the time axis; that is, the time nodes corresponding to each vector must be consistent.

[0076] The feature splicing subunit 333 splices the feature data from the six aspects mentioned above into multivariate input features, enabling the three-dimensional prediction model to simultaneously acquire information from six dimensions—generation, load, market, environment, dispatch, and behavior—at the input layer. Each dimension of the multivariate input features corresponds to a specific data source and physical meaning. These dimensions complement each other, collectively forming a complete description of the operating status of the energy storage virtual power plant. After receiving the multivariate input features, the three-dimensional prediction model can utilize the information from the six dimensions for collaborative prediction. Information from the generation and load sides provides the basis for power balance prediction; electricity price information from the market side provides signals for economic dispatch; meteorological information from the environment side provides driving conditions for new energy generation; command information from the dispatch side provides external constraints; and electricity consumption behavior information from the behavior side provides deep-seated patterns in load changes, thus achieving multi-source data-driven three-dimensional collaborative prediction.

[0077] Please see Figure 7 and Figure 8 In some implementations, the three-dimensional prediction model includes a three-layer long short-term memory network. Step 07 specifically includes: 071: inputting multivariate input features into the three-layer long short-term memory network, and the three-layer long short-term memory network synchronously outputting the predicted power generation, predicted load power, and predicted electricity price.

[0078] Correspondingly, the three-dimensional prediction model unit 33 includes a three-dimensional prediction subunit 337. The three-dimensional prediction subunit 337 has a built-in three-layer long short-term memory network and is configured to input multivariate input features into the three-layer long short-term memory network and simultaneously output the predicted power generation, predicted load power and predicted electricity price.

[0079] Long Short-Term Memory (LSTM) networks are a variant of recurrent neural networks specifically designed for processing time-series data. Compared to ordinary recurrent neural networks, LSTM networks add memory units and three gating structures to their internal structure: input gates, forget gates, and output gates. The memory units are responsible for long-term storage of historical information. The input gates control how much of the current input information is written to the memory unit, the forget gates control how much of the information from the previous time step is retained, and the output gates control how much of the information in the memory unit is output to the next layer or the final output. This gating mechanism allows LSTM networks to effectively capture long-term dependencies in time-series data while avoiding the gradient vanishing or gradient exploding problems that occur in ordinary recurrent neural networks during backpropagation. The power generation, load power, and electricity price of a virtual energy storage plant are all time-series data, and their changing patterns depend on historical states from multiple past moments. For example, load power is affected by electricity consumption trends from the previous few hours, electricity prices are affected by prices at the same time on the previous day or several days ago, and power generation is affected by weather trends from the previous few minutes. The structural characteristics of LSTM networks make them suitable for handling such time-series forecasting tasks with long-term dependencies.

[0080] A three-layer Long Short-Term Memory (LSTM) network is a deep network structure formed by stacking three LSM layers in the depth direction. The first LSM layer receives multivariate input features, performs preliminary temporal feature extraction on the input sequence, and outputs a hidden state sequence. The second LSM layer takes the hidden state sequence from the first layer as input, extracts temporal features at a higher level of abstraction, and outputs a second hidden state sequence. The third LSM layer takes the hidden state sequence from the second layer as input, further extracts high-level temporal features, and outputs the final hidden state. The advantage of this three-layer structure is that shallower networks excel at capturing local temporal patterns and short-term dependencies in the input data, while deeper networks, based on the features extracted by the previous layers, can capture more complex long-term dependencies and cross-dimensional association patterns. After multivariate input features are fed into the three-layer LSM network, the data flows sequentially through the first, second, and third layers, with each layer further abstracting and refining the temporal features of the output of the previous layer.

[0081] Synchronous output is when a three-layer long short-term memory network simultaneously outputs predicted values ​​in three dimensions at the last time step or at all time steps.

[0082] The control method and control system 100 of this application adopt a three-layer long short-term memory network to synchronously output the power generation, load power and electricity price prediction values, realize the coordinated prediction of the three dimensions, fully explore the inherent relationship between electricity price signals and power generation output and load demand, and overcome the problem of inconsistent results caused by independent prediction of each dimension.

[0083] Please see Figure 7 and Figure 16 In one embodiment, each layer of the three-layer Long Short-Term Memory (LSTM) network contains 128 hidden units, and Dropout is set to 0.3. The optimizer for the three-layer LSTM network uses the Adam algorithm with an initial learning rate of 0.001, a batch size of 32, 500 iterations, and mean squared error as the loss function. Correspondingly, the three-layer LSTM network built into the three-dimensional prediction subunit 337 is configured with the above structure and training parameters.

[0084] Hidden units are the basic computational nodes within a Long Short-Term Memory (LSTM) network layer. Each hidden unit contains a memory unit and corresponding input, forget, and output gates. The number of hidden units determines the representational power of that layer; the more hidden units, the higher the feature patterns and complexity the network can learn. Each layer contains 128 hidden units, meaning each LSM layer has 128 memory units operating in parallel. Each memory unit independently maintains its own state vector and updates it at each time step.

[0085] Dropout is a regularization technique used to prevent overfitting in neural networks. During network training, some hidden units are randomly set to zero with a certain probability, preventing them from participating in forward and backward propagation in the current iteration. Dropout randomly generates a sub-network for training in each training iteration. Because the units dropped in each iteration are different, the entire network is actually composed of many different sub-networks sharing parameters. This mechanism forces each hidden unit in the network not to overly rely on the existence of other specific units, but to independently learn useful feature representations, thereby enhancing the network's generalization ability. A dropout ratio of 0.3 means that in each training iteration, each hidden unit has a 30% probability of being randomly dropped, its activation value set to zero, and it does not participate in the calculation of this iteration. The choice of the dropout ratio needs to strike a balance between preventing overfitting and maintaining the network's representational ability. A dropout ratio of 0.3 is suitable for medium-sized time-series prediction tasks, effectively preventing a three-layer long short-term memory network from overfitting to the training data during 500 training iterations, without compromising the network's representational ability by dropping too many units.

[0086] The Adam algorithm is an adaptive moment estimation optimization algorithm widely used for parameter optimization in deep neural networks. During training, the Adam algorithm maintains two momentum terms: first-order moment estimates and second-order moment estimates. The first-order moment estimate accelerates the gradient descent process, giving parameter updates momentum inertia. The second-order moment estimate adaptively adjusts the learning rate for each parameter, allowing for differentiated update step sizes for different parameters. The Adam algorithm combines the advantages of momentum methods and adaptive learning rates, maintaining high optimization performance across different training stages and parameter dimensions. It boasts fast convergence and good robustness to hyperparameter selection. An initial learning rate of 0.001 is used, controlling the step size along the gradient direction during each parameter update. 0.001 is a widely used initial learning rate value in deep learning, ensuring stable model convergence without oscillations in the early stages of training.

[0087] Batch size is the number of samples used to calculate the gradient during each parameter update. A batch size of 32 means that 32 samples are randomly selected from the training dataset each time to form a mini-batch, the average loss on that mini-batch is calculated, and then a parameter update is performed along the gradient direction.

[0088] The number of iterations refers to the number of times the training dataset is completely traversed. Each iteration iterates through all training samples once. After 500 iterations and 500 complete traversals of the training data, the model's loss function value tends to converge, and the prediction accuracy reaches a stable state.

[0089] Mean squared error (MSE) is a commonly used loss function in regression tasks, used to measure the difference between model predictions and true values. MSE penalizes large errors more heavily because the error value is amplified after being squared, which makes the model more focused on reducing large prediction biases during training.

[0090] This application employs a three-layer Long Short-Term Memory network with 128 hidden units per layer, coupled with a Dropout ratio of 0.3, and training configurations using the Adam optimizer and mean squared error loss function. This approach effectively suppresses overfitting while ensuring the model has sufficient representational capability for complex temporal features. The 128 hidden units provide ample network capacity, the 0.3 Dropout ratio prevents overfitting, the Adam optimizer ensures efficient training convergence, and the batch size of 32 and 500 iterations enable the model to achieve stable convergence with reasonable training time and computational resource consumption. The mean squared error loss function specifically reduces large-bias predictions, ultimately achieving ideal prediction accuracy.

[0091] Please see Figure 9 and Figure 7 In one embodiment, the control method further includes: 05: Using a preset clustering algorithm, users are clustered into a preset number of categories based on their electricity consumption behavior characteristics, resulting in category labels for each user; and, 06: Embed the category labels as auxiliary features into the multivariate input features to form the updated multivariate input features.

[0092] Step 07 specifically includes: 073: Input the updated multivariate input features into the preset three-dimensional prediction model, and output the predicted power generation, predicted load, and predicted electricity price through the three-dimensional prediction model.

[0093] Correspondingly, the three-dimensional prediction model unit 33 also includes a clustering auxiliary subunit 335. The clustering auxiliary subunit 335 is configured to use a preset clustering algorithm to cluster users into a preset number of categories based on their electricity consumption behavior characteristics, obtaining category labels for each user; and to embed these category labels as auxiliary features into the multivariate input features, forming the updated multivariate input features. The three-dimensional prediction subunit 337 is further configured to input the updated multivariate input features into a three-layer long short-term memory network, simultaneously outputting predicted power generation, predicted load, and predicted electricity price.

[0094] Clustering is an unsupervised machine learning method that divides samples in a dataset into several categories based on feature similarity, ensuring that samples within the same category have high feature similarity, while samples between different categories have low feature similarity. A common clustering algorithm is K-means clustering. The core of the K-means algorithm is to minimize the sum of squared distances from all samples to the cluster centers of their respective categories, resulting in clusters that are compact within each cluster and well-spaced between clusters.

[0095] Electricity consumption behavior characteristics are feature vectors extracted from users' historical load data that reflect their electricity consumption patterns. The construction of electricity consumption behavior characteristics can include the following dimensions: daily electricity consumption curve shape characteristics, including the proportion of electricity consumption in each time period to the total electricity consumption, such as the proportion of electricity consumption in the morning, afternoon, and night; peak-valley electricity consumption characteristics, including the ratio of peak-hour electricity consumption to off-peak-hour electricity consumption, the time of maximum load occurrence, and the time of minimum load occurrence; electricity consumption stability characteristics, including the coefficient of variation of daily electricity consumption and the similarity of electricity consumption between adjacent days; and weekday and weekend electricity consumption difference characteristics, including the ratio of average weekday electricity consumption to average weekend electricity consumption and the similarity between weekday and weekend electricity consumption curves. The electricity consumption behavior of different user groups varies significantly. Factory electricity consumption is driven by production plans, with a clear peak in the electricity consumption curve during daytime working hours and a sharp drop at night and on weekends. The electricity consumption behavior of commercial complexes is affected by business hours and holidays, with significant differences in electricity consumption patterns between weekends and weekdays. Data center electricity consumption is relatively stable, maintaining a high power consumption throughout the day. Hospital electricity consumption is characterized by a uniform distribution throughout the day, with little difference between day and night. Electric vehicle charging stations have large fluctuations in electricity consumption behavior, which is affected by the temporal distribution of electric vehicle charging demand.

[0096] In one example, the preset number of categories is set to 5. The specific characteristics of the five categories of users are as follows: Category 1: Users whose electricity consumption is mainly during the day with obvious peaks and valleys, mainly including factories and office buildings. The typical characteristic of this type of user is that their electricity consumption peaks in the morning and afternoon during working hours, decreases slightly at noon, and drops significantly to a low level at night and early morning. There is a significant difference in electricity consumption between weekdays and weekends. Category 2: Users whose electricity consumption is mainly at night, mainly including data centers and nighttime manufacturing enterprises. The typical characteristic of this type of user is that their electricity consumption is higher at night and relatively lower during the day, showing a day-night distribution opposite to the regular electricity consumption pattern. Category 3: Users whose electricity consumption is evenly distributed throughout the day, mainly including hospitals. The typical characteristic of this type of user is that their electricity consumption is relatively evenly distributed throughout the day, with no significant peak-valley differences and small day-night fluctuations. Category 4: Users with special weekend electricity consumption patterns, mainly including commercial complexes. The typical characteristic of this type of user is that their electricity consumption shows a significant difference between weekdays and weekends. Weekend electricity consumption is usually higher than weekdays, and the peak and valley positions of the electricity consumption curve are different from those on weekdays. Category 5 consists of users with large fluctuations in electricity consumption and a large number of controllable loads, mainly including electric vehicle charging stations. The typical characteristics of this type of user are that their power consumption fluctuates significantly and randomly in a short period of time, the load is highly random, and the proportion of controllable load is relatively high.

[0097] Category labels are category identifiers assigned to each user after classifying them into one of the five categories mentioned above. Category labels can be represented using one-hot encoding, i.e., a five-dimensional vector representing category affiliation, where the position corresponding to the category is 1 and the rest are 0. For example, category 1 is represented as [1,0,0,0,0], category 2 as [0,1,0,0,0], and so on. Embedding category labels as auxiliary features into the multivariate input features involves concatenating the encoded vector of the category labels with the multivariate input features along the feature dimension, resulting in updated multivariate input features. The updated multivariate input features, in addition to including generation-side features, load-side features, market-side features, environmental features, dispatch-side features, and behavioral features, also include user group category information. The input of the three-dimensional prediction model includes the category affiliation of each user, enabling the model to distinguish the differences in electricity consumption patterns among different user groups.

[0098] The control method and control system 100 of this application classify users according to their electricity consumption behavior characteristics using a preset clustering algorithm, for example, into five categories. The category labels of these five user categories are then embedded as auxiliary features into the multivariate input features to participate in the prediction of the three-dimensional prediction model. Since the electricity consumption patterns of different user groups differ significantly, the load curves of users in the same category have similar variation patterns, while the load curves of different categories of users have different peak and valley positions and fluctuation characteristics. During training, the three-dimensional prediction model can learn the typical load patterns of different user categories. In actual prediction, the category of each user is determined through the input category labels, and the load power prediction output is optimized based on the typical load pattern corresponding to that category, thereby improving the accuracy of load power prediction.

[0099] Please see Figure 10 and Figure 2 In some implementations, step 09 includes: 091: When the predicted power generation is greater than the predicted load, and the current electricity price is in a low-price period, a charging control strategy is generated; 093: When the predicted power generation is less than the predicted load, and the current electricity price is during peak hours, a discharge control strategy is generated; and, 095: When the difference between the predicted power generation value and the predicted load power value is within the preset range, a standby or maintenance control strategy is generated.

[0100] Correspondingly, the edge processor 30 also includes a policy generation unit 35, which is configured to perform the above steps 091 to 095.

[0101] The strategy generation unit 35 is a functional unit inside the edge processor 30, used to generate energy storage control strategies based on the power generation prediction, load power prediction, and electricity price prediction output by the three-dimensional prediction model.

[0102] The comparison between the predicted power generation and the predicted load power reflects the power balance of the virtual power storage plant in the future. When the predicted power generation is greater than the predicted load power, it indicates that the output of new energy power generation exceeds the regional electricity demand, and the virtual power storage plant has a power surplus. In the case of a power surplus, the energy storage device 200 can choose to store the excess electricity. Storing the excess electricity and discharging it during peak electricity price periods can yield higher returns. Therefore, when the predicted power generation is greater than the predicted load power and the current electricity price is in a low-price period, the strategy generation unit 35 generates a charging control strategy, instructing the energy storage device 200 to charge according to the predicted power difference, storing the excess new energy power generation into the energy storage unit 201.

[0103] When the predicted power generation is less than the predicted load power, it indicates that the output of new energy power generation is insufficient to meet the regional electricity demand, and the energy storage virtual power plant has a power deficit. In the event of a power deficit, the energy storage device 200 can choose to release its stored electricity to make up the gap. Utilizing the energy storage device 200 to discharge and supply energy can reduce the cost of purchasing electricity from the grid. Therefore, when the predicted power generation is less than the predicted load power and the current electricity price is during peak hours, the strategy generation unit 35 generates a discharge control strategy, instructing the energy storage device 200 to discharge according to the predicted power deficit, releasing the electrical energy stored in the energy storage unit 201 to supplement the load demand.

[0104] When the difference between the predicted power generation and the predicted load power is within a preset range, it indicates that the output of new energy power generation and the demand for electricity are basically balanced, and the energy storage virtual power plant does not have a significant power surplus or deficit. At this time, the energy storage device 200 does not need to perform large-scale charging and discharging operations. The strategy generation unit 35 generates standby or maintenance control strategies, instructing the energy storage device 200 to enter standby mode or perform equipment maintenance operations during idle periods. The preset range can be set so that the difference between the predicted power generation and the predicted load power does not exceed a certain percentage of the rated power, for example, no more than five percent.

[0105] Through the aforementioned strategy, the charging and discharging decisions of the energy storage device 200 are driven by both the power balance situation and electricity price signals. It charges when both power surplus and low electricity prices are met, discharges when both power deficit and high electricity prices are met, and remains in standby mode when power balance or electricity price signals are mismatched. This ensures that the energy storage device 200 maximizes the arbitrage opportunities in electricity price signals while guaranteeing a balance between power supply and demand, thereby improving the operational economy of the energy storage virtual power plant.

[0106] Please see Figure 11 , Figure 2 and Figure 12 In some implementations, the control method further includes: 081: Determine whether the prediction deviation of any one of the predicted power generation, predicted load power, and predicted electricity price exceeds a preset deviation threshold; 083: If the deviation exceeds the preset deviation threshold, the prediction results of the other two prediction dimensions are used to perform cross-validation and reverse correction on the dimension with the prediction deviation exceeding the preset deviation threshold to obtain the corrected three-dimensional prediction value. The step of generating the energy storage control strategy is then executed using the corrected three-dimensional prediction value; and, 085: If not exceeded, then execute the steps to generate the energy storage control strategy based on the current three-dimensional predicted values.

[0107] Correspondingly, the edge processor 30 also includes a correction and reinforcement learning unit 37, which includes a deviation detection subunit 371 and a cross-validation correction subunit 373. The deviation detection subunit 371 is configured to determine whether the prediction deviation of any one of the predicted power generation, predicted load power, and predicted electricity price exceeds a preset deviation threshold. The cross-validation correction subunit 373 is configured to, when the prediction deviation of any prediction dimension exceeds the preset deviation threshold, use the prediction results of the other two prediction dimensions to perform cross-validation and reverse correction on the dimension whose prediction deviation exceeds the preset deviation threshold to obtain a corrected three-dimensional prediction value, and use the corrected three-dimensional prediction value as the final three-dimensional prediction value; when the prediction deviation does not exceed the preset deviation threshold, the current three-dimensional prediction value is used as the final three-dimensional prediction value.

[0108] After the three-dimensional prediction subunit 337 outputs the predicted power generation, load power, and electricity price, the deviation detection subunit 371 first assesses the reliability of the predicted values ​​for each dimension to determine whether there are any significant deviations from the normal range. Calculating the prediction deviation requires obtaining the actual value corresponding to the current prediction period. In real-time operation scenarios, the actual value has not yet occurred at the prediction time. Therefore, the detection of prediction deviation can be achieved in the following way: based on the statistical distribution of historical prediction errors, calculate the degree of deviation between the current predicted value and the historical value for the same period or the trend extrapolation value. When the deviation exceeds a preset deviation threshold, the reliability of the predicted value for that dimension is determined to be low.

[0109] The preset deviation thresholds include a power generation deviation threshold, a load power deviation threshold, and an electricity price deviation threshold. In one embodiment, the power generation deviation threshold is ±5%, meaning that when the deviation between the predicted power generation value and the reference value exceeds 5%, the predicted power generation value is determined to exceed the threshold. The load power deviation threshold is ±5%, meaning that when the deviation between the predicted load power value and the reference value exceeds 5%, the predicted load power value is determined to exceed the threshold. The electricity price deviation threshold is ±8%, meaning that when the deviation between the predicted electricity price value and the reference value exceeds 8%, the predicted electricity price value is determined to exceed the threshold. Power generation and load power are constrained by physical laws, and their changes are usually relatively continuous and smooth. Deviations exceeding 5% are rare under normal operating conditions; therefore, this threshold can effectively capture abnormal predicted values. Electricity prices are affected by market factors, and their fluctuation range is relatively large. The ±8% threshold takes into account the fluctuation characteristics of electricity price signals.

[0110] When electricity prices are at their peak, it typically corresponds to high load demand and a tight power grid supply. In this case, power generation should match the load capacity. Conversely, when electricity prices are at their low, it typically corresponds to low load demand and a sufficient power grid supply. These relationships can serve as constraints for cross-validation. When the prediction deviation in one dimension exceeds a preset deviation threshold, the cross-validation correction subunit 373 uses the prediction values ​​of the other two dimensions to perform a reverse correction on that dimension. The corrected three-dimensional prediction value replaces the original prediction value and is used in subsequent strategy generation steps. Through deviation detection and cross-validation correction, the three-dimensional prediction value undergoes credibility screening and anomaly correction before being input into the strategy generation unit 35, improving the reliability of the prediction results and the system robustness under abnormal scenarios.

[0111] Please continue reading. Figure 13 and Figure 12 In some implementations, the preset deviation thresholds include a power generation deviation threshold, a load power deviation threshold, and an electricity price deviation threshold. Step 083 specifically includes: 0831: When the deviation of the power generation forecast exceeds the power generation deviation threshold, the load power forecast and electricity price forecast are used for reverse correction. 0833: When the deviation of the load power forecast exceeds the load power deviation threshold, reverse correction is performed using the power generation forecast and the electricity price forecast; 0835: When the deviation of the electricity price forecast exceeds the electricity price deviation threshold, reverse correction is performed using the power generation forecast and load power forecast.

[0112] Correspondingly, the cross-validation correction subunit 373 is configured to execute steps 0831, 0833 and 0835.

[0113] When the deviation of the predicted power generation exceeds ±5%, the cross-validation correction subunit 373 uses the predicted load power and electricity price to reverse the predicted power generation. Specifically, in a power system, power generation and load power should remain balanced in steady state. The predicted load power has higher reliability than the predicted power generation, and the trend of the electricity price signal reflects the tightness of power supply and demand. The cross-validation correction subunit 373 calculates a reasonable range for power generation based on the combination of the predicted load power and the predicted electricity price, and adjusts the predicted power generation exceeding the threshold towards this reasonable range. Specifically, when the predicted electricity price is during peak hours and the predicted load power is high, it indicates strong power demand in the system, and the predicted power generation should not be too low. If the predicted power generation is abnormally low, it is adjusted towards the predicted load power. Conversely, when the predicted electricity price is during off-peak hours and the predicted load power is low, the predicted power generation should not be too high. If the predicted power generation is abnormally high, it is adjusted towards the predicted load power.

[0114] When the deviation of the load power forecast exceeds ±5%, the cross-validation correction subunit 373 uses the power generation forecast and the electricity price forecast to perform a reverse correction on the load power forecast. Specifically, load power changes usually have a certain inertia and will not change abruptly in a short period of time. The combination of the power generation forecast and the electricity price forecast can provide constraints on the reasonable range of load power. More specifically, when the power generation forecast is high and the electricity price forecast is low, it indicates that the output of new energy sources is sufficient and the electricity demand is relatively abundant, so the load power forecast should not be too low; when the power generation forecast is low and the electricity price forecast is high, it indicates that the power supply is tight and the electricity demand may decrease due to the suppression of electricity prices, so the load power forecast should not be too high.

[0115] When the deviation of the electricity price forecast exceeds ±8%, the cross-validation correction subunit 373 uses the power generation forecast and load power forecast to perform a reverse correction on the electricity price forecast. Specifically, there is an inverse relationship between electricity price and the balance of power supply and demand: electricity prices tend to be lower when there is an abundance of new energy power generation and higher when demand is strong. More specifically, when the power generation forecast is significantly greater than the load power forecast, it indicates that the power supply is abundant and the electricity price should tend to decrease; if the electricity price forecast is abnormally high, it will be adjusted downward. Conversely, when the power generation forecast is significantly less than the load power forecast, it indicates that the power supply is tight and the electricity price should tend to increase; if the electricity price forecast is abnormally low, it will be adjusted upward.

[0116] The control method and control system 100 of this application achieves accurate identification and targeted correction of dimensions exceeding the preset deviation threshold by setting differentiated deviation thresholds and corresponding cross-correction rules for the three dimensions of power generation, load, and electricity price, thereby improving the correction effect.

[0117] Please see Figure 14 and Figure 12 In one embodiment, after step 083, the control method further includes step 084: dynamically adjusting the learning rate of the 3D prediction model according to the changing trend of the prediction error. Correspondingly, the correction and reinforcement learning unit 37 further includes a learning rate adaptive adjustment subunit 375. The learning rate adaptive adjustment subunit 375 is configured to dynamically adjust the learning rate of the 3D prediction model according to the changing trend of the prediction error.

[0118] After the cross-validation correction subunit 373 completes the correction of the three-dimensional prediction value, the learning rate adaptive adjustment subunit 375 collects the prediction error data of the current prediction period, analyzes the trend of the prediction error, and dynamically adjusts the learning rate parameter used by the three-dimensional prediction model in subsequent training or fine-tuning.

[0119] Step 084 specifically includes: 0841: Obtain the generation forecast error, load forecast error, and electricity price forecast error for the current forecast period. The length of the forecast period can be determined according to the system configuration, for example, one forecast period per hour or one forecast period per 15 minutes. The generation forecast error is the deviation between the predicted generation power and the actual generation power; the load forecast error is the deviation between the predicted load power and the actual load power; and the electricity price forecast error is the deviation between the predicted electricity price and the actual electricity price.

[0120] 0843: Calculate the sum of the generation forecast error, load forecast error, and electricity price forecast error as the total error. The total error is calculated as the sum of the absolute values ​​or squares of the forecast errors in the three dimensions, or as a weighted sum of the forecast errors in the three dimensions. The weights of different dimensions can be determined based on the importance of each dimension in the energy storage control strategy. The total error reflects the overall forecast accuracy of the three-dimensional forecast model within the current forecast period.

[0121] 0845: When the total error shows an increasing trend for a preset number (e.g., 3) of prediction periods, increase the learning rate of the 3D prediction model by a first preset percentage (e.g., 10%). The determination of an increasing trend in the total error over 3 consecutive prediction periods is as follows: compare the total error values ​​of the most recent three consecutive prediction periods. If the total error of each period is greater than the total error of the previous period, then the total error is considered to be increasing. A continuous increase in the total error indicates a continuous decline in the prediction accuracy of the 3D prediction model. In this case, increasing the learning rate by 10% allows the model to adjust parameters with larger step sizes in subsequent training, accelerating the model's response to error changes.

[0122] 0847: When the total error of a predetermined number of prediction periods (e.g., 5) shows a decreasing trend and tends to stabilize, the learning rate is reduced by a second preset percentage (e.g., 5%). The determination that the total error of 5 consecutive prediction periods shows a decreasing trend and tends to stabilize is made by comparing the total error values ​​of the most recent five prediction periods. If the total error of each period is less than the total error of the previous period, and the difference between the total errors of the most recent two periods is less than a certain preset stabilization threshold, then the total error is determined to be decreasing and tending to stabilize. A continuously decreasing and stabilizing total error indicates that the 3D prediction model has good prediction accuracy and a stable state. At this point, the learning rate is reduced by 5%, allowing the model to fine-tune parameters with smaller step sizes in subsequent training, avoiding parameter oscillations and overfitting.

[0123] The learning rate is adjusted within the range of 0.001 to 0.01. 0.001 is the lower limit of the learning rate; once it's reached, it will not be adjusted downwards. 0.01 is the upper limit of the learning rate; once it's reached, it will not be adjusted upwards. Fluctuating the learning rate within this range ensures the model has sufficient learning capacity while avoiding instability caused by an excessively large learning rate or slow convergence caused by an excessively small learning rate.

[0124] This application judges the model performance trend by the cumulative prediction of the total error, and adjusts the learning rate differently when the error continues to increase or decreases and tends to stabilize, thus achieving precise and stable adjustment of the learning rate.

[0125] The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as the combination of these technical features does not contradict each other, it should be considered within the scope of this specification. At the same time, other implementation methods can be derived from the above embodiments, so that structural and logical substitutions and changes can be made without departing from the scope of this disclosure.

[0126] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A control method for an energy storage virtual power plant, wherein the energy storage virtual power plant includes energy storage equipment, characterized in that, include: Acquire multi-source data, including photovoltaic power generation data, wind power generation data, load power consumption data, time-of-use electricity price data, meteorological data, power grid dispatch instructions, and user electricity consumption behavior data; The multi-source data is preprocessed to obtain standardized time-series data; The photovoltaic power generation data and wind power generation data in the standardized time series data are weighted and fused using preset attention weights to obtain fused feature data; The fused feature data is combined with the load electricity consumption data, time-of-use electricity price data, meteorological data, power grid dispatch instructions and user electricity consumption behavior data in the standardized time series data to form multivariate input features; The multivariate input features are input into a preset three-dimensional prediction model, and the three-dimensional prediction model outputs the predicted power generation, predicted load, and predicted electricity price. An energy storage control strategy is generated based on the predicted power generation, the predicted load power, and the predicted electricity price. and The operating mode of the energy storage device is controlled according to the energy storage control strategy.

2. The control method according to claim 1, characterized in that, The acquisition of multi-source data includes: The photovoltaic power generation data is collected by the photovoltaic inverter at a first preset acquisition frequency; The wind power generation data is collected by the wind power converter at a second preset acquisition frequency; The load power consumption data is collected by a smart meter at a third preset collection frequency; The time-of-use electricity price data is acquired through the power grid data platform at a fourth preset acquisition frequency. The meteorological data is acquired through a weather station or a meteorological data service provider at the fifth preset acquisition frequency; The power grid dispatching instructions are obtained through the power grid energy management system; The user's electricity consumption data is collected through smart meters.

3. The control method according to claim 1, characterized in that, The preprocessing of the multi-source data includes: The multi-source data from different acquisition frequencies are uniformly resampled to the same preset reference frequency to align the multi-source data on the time axis, so as to obtain aligned data. Abnormal data points in the aligned data are removed based on preset criteria to obtain the first intermediate data; For data in the first intermediate data whose consecutive missing time does not exceed a preset time threshold, linear interpolation is used to complete the data; for data whose consecutive missing time exceeds the preset time threshold, historical data from the same period is used to replace the missing data to obtain the second intermediate data; and, The second intermediate data is normalized using a preset standardization method to obtain the standardized time series data.

4. The control method according to claim 1, characterized in that, The meteorological factors in the meteorological data include light intensity, wind speed, and temperature; the weighted fusion of photovoltaic power generation data and wind power generation data in the standardized time-series data using preset attention weights includes: The preset attention weights are calculated using the Softmax function. These preset attention weights include light intensity weights, wind speed weights, and temperature weights, and the sum of these weights is 1. The photovoltaic power generation data and wind power generation data at corresponding times in the standardized time series data are weighted using the light intensity weight, the wind speed weight, and the temperature weight, respectively, to obtain the fused feature data.

5. The control method according to claim 1, characterized in that, The three-dimensional prediction model includes a three-layer long short-term memory network. The process of inputting the multivariate input features into the preset three-dimensional prediction model and outputting predicted power generation, predicted load, and predicted electricity price through the three-dimensional prediction model includes: The multivariate input features are input into the three-layer long short-term memory network, and the three-layer long short-term memory network synchronously outputs the predicted power generation value, the predicted load power value, and the predicted electricity price value.

6. The control method according to claim 5, characterized in that, In the three-layer long short-term memory network, each layer contains 128 hidden units, and Dropout is set to 0.3; The optimizer of the three-layer long short-term memory network uses the Adam algorithm, with an initial learning rate of 0.001, a batch size of 32, 500 iterations, and mean squared error as the loss function.

7. The control method according to claim 1, characterized in that, Also includes: A preset clustering algorithm is used to cluster users into a preset number of categories based on their electricity consumption behavior characteristics, thereby obtaining the category label for each user; and, The category label is embedded as an auxiliary feature into the multivariate input feature to form the updated multivariate input feature; The step of inputting the multivariate input features into a preset three-dimensional prediction model, and outputting power generation prediction, load prediction, and electricity price prediction values ​​through the three-dimensional prediction model, includes: The updated multivariate input features are input into a preset three-dimensional prediction model, and the three-dimensional prediction model outputs the predicted power generation, predicted load, and predicted electricity price.

8. The control method according to claim 1, characterized in that, The generation of an energy storage control strategy based on the predicted power generation, the predicted load power, and the predicted electricity price includes: When the predicted power generation value is greater than the predicted load power value, and the current electricity price is in a low-off period, a charging control strategy is generated. When the predicted power generation is less than the predicted load, and the current electricity price is during peak hours, a discharge control strategy is generated; and, When the difference between the predicted power generation value and the predicted load power value is within a preset range, a standby or maintenance control strategy is generated.

9. The control method according to claim 1, characterized in that, Also includes: Determine whether the prediction deviation of any one of the predicted power generation value, the predicted load power value, and the predicted electricity price value exceeds a preset deviation threshold. If the deviation exceeds the preset deviation threshold, the prediction results of the other two prediction dimensions are used to perform cross-validation and reverse correction on the dimension whose prediction deviation exceeds the preset deviation threshold to obtain the corrected three-dimensional prediction value, and the step of generating an energy storage control strategy is executed with the corrected three-dimensional prediction value. If the value is not exceeded, then the steps for generating the energy storage control strategy will be executed based on the current three-dimensional predicted value.

10. The control method according to claim 9, characterized in that, The preset deviation thresholds include: a power generation deviation threshold, a load power deviation threshold, and an electricity price deviation threshold; the step of using the prediction results of the other two prediction dimensions to perform cross-validation and reverse correction on the dimensions whose prediction deviation exceeds the preset deviation thresholds to obtain corrected three-dimensional prediction values ​​includes: When the deviation of the predicted power generation exceeds the predicted power generation deviation threshold, the predicted load power and the predicted electricity price are used for reverse correction. When the deviation of the predicted load power exceeds the predicted load power deviation threshold, the predicted power generation and the predicted electricity price are used for reverse correction. When the deviation of the predicted electricity price exceeds the predicted electricity price deviation threshold, the predicted power generation and the predicted load power are used for reverse correction.

11. The control method according to claim 9, characterized in that, After performing cross-validation and reverse correction on the dimension whose prediction deviation exceeds the preset deviation threshold using the prediction results of the other two prediction dimensions to obtain the corrected three-dimensional prediction value, the control method further includes: The learning rate of the three-dimensional prediction model is dynamically adjusted based on the changing trend of the prediction error using a preset reinforcement learning method.

12. The control method according to claim 11, characterized in that, The step of dynamically adjusting the learning rate of the 3D prediction model using a preset reinforcement learning method based on the changing trend of the prediction error includes: Obtain the generation forecast error, load forecast error, and electricity price forecast error for the current forecast period; The sum of the power generation forecast error, the load forecast error, and the electricity price forecast error is calculated as the total error; When the total error shows an increasing trend over a consecutive preset number of prediction periods, the learning rate of the three-dimensional prediction model is increased by a first preset percentage; and, When the total error over a predetermined number of prediction periods shows a decreasing trend and tends to stabilize, the learning rate is reduced by a second preset percentage; the adjustment range of the learning rate is controlled between 0.001 and 0.

01.

13. A control system for an energy storage virtual power plant, characterized in that, include: The acquisition module is configured to acquire data from multiple sources. An edge processor, connected to the acquisition module, is configured to execute the control method according to any one of claims 1-12 to generate an energy storage control strategy; and, An energy storage control module is connected to the edge processor and the energy storage device respectively, and is configured to: receive the energy storage control strategy generated by the edge processor, and control the working mode of the energy storage device according to the energy storage control strategy.

Images