Virtual power plant regulation capability dynamic prediction system
By constructing a multi-source heterogeneous data real-time processing platform and a dual-channel deep learning model, the accuracy and real-time performance issues of the virtual power plant prediction system were solved, achieving high-precision, adaptive regulation capability prediction and supporting the rapid development of virtual power plants.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-02
- Publication Date
- 2026-03-13
AI Technical Summary
Existing virtual power plant forecasting systems lack sophisticated modeling of the dynamic coupling and mutual constraints between heterogeneous resources, resulting in low forecast accuracy, inability to adapt to changes in the external environment, and difficulty in meeting high real-time requirements with cloud computing technology, leading to high system expansion costs.
We construct a real-time processing platform that integrates multi-source heterogeneous data, design a dedicated dual-channel deep learning prediction model, and combine it with edge gateways and cloud platforms to achieve real-time data processing and online self-learning, thereby improving prediction accuracy and adaptability.
It improves the accuracy and real-time performance of virtual power plant regulation capacity prediction, enables it to adapt to dynamic changes, provides reliable basis for power grid dispatching decisions, and enhances operational economy and power grid support capabilities.
Smart Images
Figure CN121660163A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of power systems and energy management, and specifically relates to a dynamic prediction system for the regulation capacity of a virtual power plant. Background Technology
[0002] Early virtual power plant solutions mostly adopted centralized data management based on cloud platforms. Their prediction methods usually relied on historical processing data and load data of distributed resources (such as distributed photovoltaics and energy storage) within the virtual power plant. They used time series analysis or traditional machine learning algorithms to predict the regulation capacity for the next few hours. This method treated the virtual power plant as a whole and extrapolated based on historical patterns. With the improved accessibility of meteorological data, researchers have begun to incorporate meteorological forecast data into prediction models, especially for the prediction of photovoltaic power output. By establishing statistical relationships between irradiance, temperature and photovoltaic power output, the prediction results are corrected, improving the prediction accuracy of the photovoltaic power plant's own output and indirectly affecting the assessment of the virtual power plant's regulation capability. There are also independent prediction systems for single types of distributed resources (such as energy storage clusters and electric vehicle charging stations). These systems focus on the operational characteristics of that type of distributed resource, but they are usually isolated and have not yet been deeply coupled and collaboratively modeled with other types of distributed resources in the virtual power plant.
[0003] Cloud service platforms from companies like Microsoft, Google, and IBM provide services such as virtual machines, storage, databases, artificial intelligence, and machine learning. While the development and application of cloud computing technology have provided high-quality services, issues remain regarding security and data privacy. Virtual power plant operators in Europe and North America incorporate economic factors such as electricity price signals and ancillary service market clearing prices into their forecasts. By combining real-time electricity price forecasts with assessments of the willingness of adjustable loads to respond at different price levels, they can dynamically adjust the range of adjustable capacity they can provide. This method is relatively well-developed in terms of economic models, but it is somewhat lacking in real-time perception of the physical characteristics of resources. Lawrence Berkeley National Laboratory in the United States has also explored agent-based modeling for virtual power plant simulation. This involves establishing an agent model for each distributed resource to simulate its behavior under different external conditions. The overall characteristics of the virtual power plant emerge through the interaction of numerous agents. This method is mainly used for policy effect simulation and operational strategy research, but due to its high computational complexity, it has not yet been widely applied to online real-time forecasting.
[0004] In summary, existing technologies, whether treating the virtual power plant as a whole or making independent predictions for a single resource type, lack refined modeling of the dynamic coupling and mutual constraints between heterogeneous resources. This fails to accurately reflect the synergistic or bottleneck effects generated after resource aggregation, leading to significant prediction errors in aggregation regulation capabilities. Existing models rely too heavily on historical power data or market signals, failing to achieve deep integration of multi-dimensional data. When the resource composition of the virtual power plant changes dynamically or the external environment fluctuates drastically, static models cannot adaptively adjust, resulting in a sharp decline in prediction accuracy. While cloud computing technology offers advanced models, as a general-purpose solution, its models do not embed domain knowledge of the power system and virtual power plants, and the long data processing chain makes it difficult to meet high real-time requirements. Closed systems customized for specific scenarios lack a standardized, pluggable prediction framework to support the rapid access of new distributed resources and agile iteration of prediction models, resulting in high system expansion costs and long cycles, making it difficult to adapt to the rapid development of virtual power plant business. Summary of the Invention
[0005] To address the problems in the prior art, this invention proposes a dynamic prediction system for the regulation capacity of a virtual power plant. By constructing a real-time processing platform that integrates multi-source heterogeneous data and designing a dedicated dual-channel deep learning prediction model, it effectively overcomes the shortcomings of the prior art in terms of prediction accuracy, real-time performance, and adaptability.
[0006] The technical solution of the present invention is as follows: This invention provides a dynamic prediction system for the regulation capacity of a virtual power plant, comprising a data acquisition and edge processing layer, a cloud platform data processing and storage layer, an intelligent prediction model layer, an application and service platform layer, a user interaction layer, and a system management layer; The data acquisition and edge processing layer includes a sensing and control unit and an edge gateway deployed at the distributed resource aggregation point. The sensing and control unit collects real-time operating parameter data of the distributed resources; the edge gateway preprocesses the real-time operating parameter data. The cloud platform's data processing and storage layer performs feature engineering on preprocessed data through stream processing and batch processing, and stores the preprocessed data in the database; The intelligent prediction model layer trains the intelligent prediction model through feature-engineered data, predicts the regulation capacity of virtual power plants, and performs version management, deployment, and online updates of the intelligent prediction model; The application and service platform layer displays the prediction results and provides query services; The user interaction layer manages and operates the prediction system, providing users with operational tools; The system management layer monitors and manages the data in the system.
[0007] Furthermore, the sensing and control unit includes a smart meter, an inverter controller, an energy storage management system, and a load controller; The real-time operating parameters include active power, reactive power, switching status, and state of charge.
[0008] Furthermore, the preprocessing specifically involves performing real-time operating parameter data protocol conversion, and performing preliminary data cleaning and caching.
[0009] Furthermore, the cloud platform data processing and storage layer adopts a distributed stream processing platform to receive data streams from the edge gateway, perform streaming and batch processing, and store the data; The stored data includes a time-series database and a relational database. The time-series database is used to store the raw data after streaming processing; the relational database is used to store distributed resource metadata, model parameters, and prediction results.
[0010] Furthermore, the streaming processing employs a distributed real-time streaming framework for real-time data cleaning, outlier detection, and data alignment. The batch processing uses a distributed big data computing framework to perform feature engineering on historical data.
[0011] Furthermore, the intelligent prediction model layer includes a feature extraction module, a dual-channel deep learning module, and a model management module. The feature extraction module extracts key features from the raw data after streaming processing, including historical features, external features, and state features. The dual-channel deep learning module includes a resource time-series channel and an external feature channel. The resource time-series channel processes historical feature data, and the external feature channel processes external feature data. The output vectors of the resource time-series channel and the external feature channel are concatenated and passed through a preset number of fully connected layers to output a predicted value of the adjustment capability. The model management module is responsible for the version management, deployment, and online updates of the intelligent prediction model.
[0012] Furthermore, the historical characteristics are historical power storage and time-series statistical characteristics, and the time-series statistical characteristics include the mean and variance of historical power. The external features include forecast data for future temperature, humidity, wind speed, and irradiance, as well as future electricity price curves; The state characteristics include the energy storage system's state of charge and equipment health indicators.
[0013] Furthermore, the application and service platform layer displays historical data, prediction results, and prediction errors in the form of curves and charts through a visual human-computer interface; The application and service platform layer also includes a system management console, which provides model parameter configuration, resource management and user permission management functions, and provides prediction result query services through the representative state transition application programming interface box, virtual power plant energy management system and dispatch center.
[0014] Furthermore, the user interaction layer performs management operations through a web interface or application programming interface provided by the management console. These management operations include resource allocation, application deployment, and monitoring and alarms. The user interaction layer provides users with command-line tools or graphical interface tools through client tools for daily operation and maintenance.
[0015] Furthermore, the monitoring and management strategies include data quality management strategies, model lifecycle management strategies, resource management and scheduling strategies, and security and privacy management strategies.
[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention overcomes the shortcomings of existing technologies in prediction accuracy, real-time performance, and adaptability by constructing a real-time processing platform that integrates multi-source heterogeneous data and designing a dedicated dual-channel deep learning prediction model. Specifically, the model's explicit modeling of external dynamic factors and extraction of historical time-series deep features jointly ensure high prediction accuracy; the collaborative architecture of cloud platform and edge gateway ensures low latency in data processing; and the online self-learning mechanism enables the system to adapt to the dynamic changes of virtual power plants and maintain long-term prediction accuracy. Therefore, this invention can provide more reliable decision-making basis for virtual power plants to participate in grid dispatch and market transactions, significantly improving the economic efficiency of virtual power plant operation and its support capability for the power grid. Attached Figure Description
[0017] Figure 1 The operation and scheduling process of the virtual power plant regulation capacity dynamic prediction system. Detailed Implementation
[0018] To make the features and beneficial effects of the present invention more apparent and understandable, the present invention will be further described below in conjunction with the accompanying drawings and embodiments.
[0019] Example 1 This embodiment provides a virtual power plant regulation capacity dynamic prediction system, including a data acquisition and edge processing layer, a cloud platform data processing and storage layer, an intelligent prediction model layer, an application and service platform layer, a user interaction layer, and a system management layer; The data acquisition and edge processing layer includes a sensing and control unit and an edge gateway deployed at the distributed resource aggregation point. The sensing and control unit includes smart meters, inverter controllers, energy storage management systems, and load controllers, and is responsible for collecting real-time operating parameters of various distributed resources, such as active power, reactive power, switching status, and state of charge. The edge gateway is responsible for converting the Modicon protocol or object-oriented asynchronous service access point protocol into a message queue telemetry transmission protocol, performing preliminary data cleaning and caching on the real-time operating parameters of distributed resources, and transmitting the preliminarily cleaned data to the cloud platform. The cloud platform's data processing and storage layer uses a high-concurrency message middleware (such as the distributed stream processing platform Kafka) to receive data streams from edge gateways, perform streaming and batch processing, and store the data. The streaming processing uses the distributed real-time stream processing framework Apache Flink for real-time data cleaning, outlier detection, and data alignment. The batch processing uses the distributed big data computing framework Apache Spark to perform feature engineering on historical data. The stored data includes a time-series database and a relational database. The time-series database is used to store the raw data after streaming processing, and the relational database is used to store distributed resource metadata, model parameters, and prediction results. The intelligent prediction model layer includes a feature extraction module, a dual-channel deep learning module, and a model management module. The feature extraction module extracts key features from the raw data after streaming processing, including historical features, external features, and state features. The historical features are historical power sequences and time-series statistical features, including the mean and variance of historical power. The external features include forecast data for future temperature, humidity, wind speed, and irradiance, as well as future electricity price curves. The state features include the state of charge of the energy storage system and equipment health indicators. The dual-channel deep learning module includes a resource temporal channel and an external feature channel. The input to the resource temporal channel is historical feature temporal data, which is sequentially passed through a one-dimensional convolutional layer to extract local fluctuation patterns, a gated recurrent unit layer to capture long-term temporal dependencies, and an attention layer to calculate the importance weights of different historical moments. The input to the external feature channel is data-aligned external feature temporal data, which is then processed by a fully connected layer for feature encoding. The output vectors of the resource temporal channel and the external feature channel are concatenated and then passed through a preset number of fully connected layers to output the predicted values of the maximum up-adjustment capability and the maximum down-adjustment capability at future time points. The model management module is responsible for the version management, deployment, and online updates of the intelligent prediction model; The application and service platform layer displays historical data, prediction results, and prediction errors in the form of curves and charts through a visual human-computer interface. The system management console provides model parameter configuration, resource management, and user permission management functions, and provides prediction result query services to the virtual power plant energy management system and dispatch center through the representative state transition (REST) application programming interface (API). The user interaction layer includes a management console and client tools. The management console provides a web interface or application programming interface for management operations, including resource allocation, application deployment, and monitoring and alarms. The client tools provide users with command-line tools or graphical interface tools for daily operation and maintenance. The system management layer monitors and manages the data in the system. The management strategies include data quality management strategy, model lifecycle management strategy, resource management and scheduling strategy, and security and privacy management strategy.
[0020] Preferably, the data quality management strategy performs data credibility assessment and data consistency verification on the real-time operating parameters of various distributed resources. The data credibility assessment configures a credibility weight for the data source of the real-time operating parameters of each distributed resource. For devices with communication interruptions or data anomalies, the weight of their data in the features is dynamically reduced or a difference algorithm is used to supplement it. The data consistency verification specifically involves periodically verifying the consistency of data from different sources and triggering an alarm when a deviation is detected. The model lifecycle management strategy involves conducting parallel A / B testing on the new and old models before the intelligent prediction model is updated and launched. Once the performance improvement of the new model is confirmed, the model is switched to the old model. At the same time, historical versions of the old model are saved, and when the new model has prediction problems, it is rolled back to a stable historical version. The resource management and scheduling strategy establishes a dynamic profile for each distributed resource in the access system, records the response characteristics, reliability, and adjustment potential of the distributed resources, and serves as a supplement to feature engineering; it also adds a confidence score to each prediction result to provide decision-making reference for schedulers. The security and privacy management strategy includes data transmission encryption and access control. The data transmission encryption uses transport layer security protocols to encrypt data transmission between the edge gateway and the cloud platform and between services within the cloud platform. The access control implements role-based access control, allowing only authorized users and authorized applications to access distributed resource data and intelligent prediction model services.
[0021] Example 2 This embodiment provides an operation and scheduling process for a virtual power plant dynamic prediction system for regulation capacity, such as... Figure 1 As shown, the scheduling process includes the following steps: When the system starts, it loads the latest intelligent prediction model from the historical model version library, starts the data access service, and establishes a stable data connection between the edge gateway and the cloud platform. The edge gateway continuously uploads real-time operating parameter data of various distributed resources to the cloud platform message queue. Stream processing is used to parse the data in real time and filter out outliers. The data is then aligned and aggregated according to a unified time window to form the input features required by the intelligent prediction model. A prediction task is automatically triggered whenever the preset period is reached. The intelligent prediction model layer obtains the latest feature data from the time series database and relational database, inputs it into the intelligent prediction model for forward calculation, and outputs the upper and lower limit prediction curves of the virtual power plant regulation capacity within the prediction time range. The prediction results are published to the upper-layer applications that subscribe to the prediction service through the application programming interface, and the prediction results and corresponding feature data are stored in the database for subsequent model evaluation and retraining. The system management layer continuously compares predicted values with actual measured values, calculates mean absolute error and root mean square error indices. When the performance indicators continue to deteriorate or reach the preset retraining cycle, the offline retraining process of the entire prediction model is automatically triggered. The feature data is updated using a new database. After the retraining performance indicators are verified, the new model is seamlessly switched to the online service environment to replace the old model.
[0022] Example 3 This embodiment provides an implementation process for deploying a virtual power plant regulation capacity dynamic prediction system in a virtual power plant, including the following steps: Define the time scale and accuracy requirements for forecasting, deploy edge gateways, and complete the communication integration of various distributed resources; Deploy various middleware and databases required for cloud platform data processing and storage; Based on historical data from distributed resources, feature engineering is performed to design and train a dual-channel deep learning model. Integrate all modules of the intelligent prediction model layer and deploy cloud services; The system was tested in a real power environment, and the parameters of the dual-channel deep learning model and the system configuration were adjusted based on the feedback from the test results. After trial operation and adjustments, the system was officially put into operation as a virtual power plant, and continuous monitoring, maintenance and dual-channel deep learning model iterative updates were performed.
Claims
1. A virtual power plant regulation capacity dynamic prediction system, characterized in that, It includes a data acquisition and edge processing layer, a cloud platform data processing and storage layer, an intelligent prediction model layer, an application and service platform layer, a user interaction layer, and a system management layer; The data acquisition and edge processing layer includes a sensing and control unit and an edge gateway deployed at the distributed resource aggregation point. The sensing and control unit collects real-time operating parameter data of the distributed resources; the edge gateway preprocesses the real-time operating parameter data. The cloud platform's data processing and storage layer performs feature engineering on preprocessed data through stream processing and batch processing, and stores the preprocessed data in the database; The intelligent prediction model layer trains the intelligent prediction model through feature-engineered data, predicts the regulation capacity of virtual power plants, and performs version management, deployment, and online updates of the intelligent prediction model; The application and service platform layer displays the prediction results and provides query services; The user interaction layer manages and operates the prediction system, providing users with operational tools; The system management layer monitors and manages the data in the system.
2. The virtual power plant regulation capacity dynamic prediction system according to claim 1, characterized in that, The sensing and control unit includes a smart meter, an inverter controller, an energy storage management system, and a load controller. The real-time operating parameters include active power, reactive power, switching status, and state of charge.
3. The virtual power plant regulation capacity dynamic prediction system according to claim 1, characterized in that, The preprocessing specifically involves performing real-time operating parameter data protocol conversion, and performing preliminary data cleaning and caching.
4. The virtual power plant regulation capacity dynamic prediction system according to claim 1, characterized in that, The cloud platform data processing and storage layer uses a distributed stream processing platform to receive data streams from the edge gateway, perform streaming and batch processing, and store the data. The stored data includes a time-series database and a relational database. The time-series database is used to store the raw data after streaming processing; the relational database is used to store distributed resource metadata, model parameters, and prediction results.
5. The virtual power plant regulation capacity dynamic prediction system according to claim 4, characterized in that, The streaming process employs a distributed real-time streaming framework for real-time data cleaning, outlier detection, and data alignment. The batch processing uses a distributed big data computing framework to perform feature engineering on historical data.
6. The virtual power plant regulation capacity dynamic prediction system according to claim 1, characterized in that, The intelligent prediction model layer includes a feature extraction module, a dual-channel deep learning module, and a model management module. The feature extraction module extracts key features from the raw data after streaming processing, including historical features, external features, and state features. The dual-channel deep learning module includes a resource time-series channel and an external feature channel. The resource time-series channel processes historical feature data, and the external feature channel processes external feature data. The output vectors of the resource time-series channel and the external feature channel are concatenated and passed through a preset number of fully connected layers to output a predicted value of the adjustment capability. The model management module is responsible for the version management, deployment, and online updates of the intelligent prediction model.
7. The virtual power plant regulation capacity dynamic prediction system according to claim 6, characterized in that, The historical characteristics are historical power storage and time-series statistical characteristics, including the mean and variance of historical power. The external features include forecast data for future temperature, humidity, wind speed, and irradiance, as well as future electricity price curves; The state characteristics include the energy storage system's state of charge and equipment health indicators.
8. The virtual power plant regulation capacity dynamic prediction system according to claim 1, characterized in that, The application and service platform layer displays historical data, prediction results, and prediction errors in the form of curves and charts through a visual human-computer interface. The application and service platform layer also includes a system management console, which provides model parameter configuration, resource management and user permission management functions, and provides prediction result query services through the representative state transition application programming interface box, virtual power plant energy management system and dispatch center.
9. The virtual power plant regulation capacity dynamic prediction system according to claim 1, characterized in that, The user interaction layer performs management operations through a web interface or application programming interface provided by the management console. The management operations include resource allocation, application deployment, and monitoring and alarms. The user interaction layer provides users with command-line tools or graphical interface tools through client tools for daily operation and maintenance.
10. The virtual power plant regulation capacity dynamic prediction system according to claim 1, characterized in that, The monitoring and management strategies include data quality management strategies, model lifecycle management strategies, resource management and scheduling strategies, and security and privacy management strategies.