Virtual power plant scheduling method and system based on regional power utilization activity prediction

Abstract

The invention relates to the technical field of power system management, in particular to a virtual power plant scheduling method and system based on regional power utilization activity prediction. The method comprises the following steps: collecting user load data and geographic information data of a target area; dividing users into different power utilization activity types based on the geographic information data and the user load data; for each power utilization activity type, utilizing a preset machine learning sub-model to predict a power change curve of the power utilization activity of the type in a future time period; and solving an optimal control sequence of a future time period based on a preset optimization target by using the power utilization activity prediction result, and generating a scheduling instruction for distributed energy, an energy storage system or a controllable load in the virtual power plant. According to the method, load prediction is converted into activity prediction, so that the scheduling decision of the virtual power plant is not based on a fuzzy aggregation value any more, but based on a clear portrait of a future power consumption behavior, and the scheduling precision and stability of the virtual power plant are improved.

Description

A Virtual Power Plant Dispatch Method and System Based on Regional Electricity Consumption Forecasting Technical Field

[0001] This invention relates to the field of power system management technology, specifically to a virtual power plant dispatching method and system based on regional electricity consumption activity forecasting. Background Technology

[0002] As the global energy transition deepens, the form of the power system is evolving from traditional centralized power generation to a decentralized, flattened structure. To address the intermittency and volatility of renewable energy sources (such as solar and wind power), virtual power plants have emerged as a solution. Virtual power plants do not rely on single large-scale generator units; instead, they aggregate geographically dispersed and diverse distributed energy resources into a unified and coordinated entity through software, control, and communication technologies. These distributed energy resources can include distributed power sources (such as solar and wind power), energy storage systems, controllable loads, and electric vehicles and their charging stations. As a whole, virtual power plants can participate in the electricity market, provide grid ancillary services (such as frequency regulation and reserve), thereby improving the utilization rate of renewable energy, enhancing grid stability, and creating economic value for participants.

[0003] To achieve the economical and stable operation of virtual power plants, the core lies in optimizing the scheduling strategy. The scheduling of virtual power plants is typically constructed as an optimization problem, aiming to maximize the overall revenue or minimize the operating costs of the virtual power plant while satisfying various operational constraints. These optimization models can be deterministic, stochastic optimization considering uncertainty, or robust optimization.

[0004] However, the effectiveness of all these optimized scheduling methods heavily relies on one of their key inputs: accurate forecasting of future electricity demand. In existing technologies, this is often referred to as "load forecasting." Traditional load forecasting methods, whether based on time-series statistical models (such as ARIMA) or artificial intelligence-based models (such as neural networks), are essentially regression tasks. They attempt to predict an aggregate, macroscopic load value (e.g., the total electricity consumption of a region in the next hour).

[0005] Existing load forecasting technologies suffer from the following problems: First, they blur the diverse electricity consumption behaviors of all users within a region into a single curve, thus losing crucial granular information. Second, they fail to distinguish between load types or causes. For example, a 10MW load peak might be caused by 1000 electric vehicles charging simultaneously (a rigid, high-power load) or by 5000 households simultaneously turning on their air conditioners (a load with thermal inertia and flexible dispatchability). For virtual power plant dispatching, the "dispatchability potential" and "response strategies" for these two scenarios are drastically different, but traditional load forecasting models cannot provide this distinction.

[0006] With the rapid growth of new types of loads (such as electric vehicles and data centers), the highly variable and uncertain characteristics of these loads cause a sharp increase in the error of traditional aggregated forecasting. The optimal scheduling model of a virtual power plant receives such low-fidelity, opaque, and highly uncertain load forecasts, and its scheduling decisions (such as energy storage charging and discharging, and controllable load interruption) are inevitably suboptimal, leading to a decline in the economic benefits of virtual power plants and an increase in grid management risks. Summary of the Invention

[0007] To address the shortcomings of existing technologies, this invention provides a virtual power plant scheduling method and system based on regional electricity consumption activity prediction.

[0008] A virtual power plant scheduling method based on regional electricity consumption activity prediction includes the following steps: collecting user load data, geographic information data, and activity-related data including electric vehicle connection status and industrial production plans in the target area; constructing an electricity consumption activity type profile containing typical electricity consumption behavior characteristics based on the geographic information data and user load data, and classifying users into different electricity consumption activity types; predicting the power change curve of the type of electricity consumption activity in future periods using a preset machine learning sub-model for each electricity consumption activity type; using the electricity consumption activity prediction results, solving for the optimal control sequence in future periods based on a preset optimization objective, and generating scheduling instructions for distributed energy, energy storage systems, or controllable loads in the virtual power plant, wherein the optimized scheduling model is a model predictive control model.

[0009] Preferably, the step of constructing a profile of electricity consumption activity types containing typical electricity consumption behavior characteristics based on the geographic information data and user load data, and classifying users into different electricity consumption activity types, specifically includes: dividing grid users into multiple topologically independent power supply areas on the distribution network based on the geographic information data; obtaining the user electricity consumption matrix of each power supply area, and separating the user feature matrix from the user electricity consumption matrix using a joint matrix tri-decomposition algorithm; and clustering multiple groups of electricity consumption activity types with similar electricity consumption behaviors based on the user feature matrix using a fuzzy C-means algorithm optimized and improved by Grey Wolf.

[0010] Preferably, obtaining the user electricity consumption matrix for each power supply area specifically includes: acquiring the original load data of the power supply area at fixed time steps, the original load data including active power, reactive power, voltage, current and corresponding timestamps; performing time alignment, missing value processing, outlier removal and normalization preprocessing on the original load data; and constructing a user electricity consumption data matrix based on the preprocessed data. In the formula, Indicates the first A power consumption data matrix for each power supply area, containing elements of the matrix. For the first The user in the first Normalized load values ​​at each time step.

[0011] Preferably, the machine learning sub-model includes an electric vehicle activity prediction sub-model and an industrial production activity prediction sub-model; the electric vehicle activity prediction sub-model uses variational mode decomposition combined with a bidirectional long short-term memory network to receive historical charging data and traffic flow data, and outputs an electric vehicle charging power change curve; the industrial production activity prediction sub-model uses an ensemble learning algorithm to receive production plan data and output an industrial base load curve.

[0012] Preferably, the step of solving for the optimal control sequence for future time periods based on a preset optimization objective and generating scheduling instructions for distributed energy sources, energy storage systems, or controllable loads in the virtual power plant specifically includes: at each scheduling time... Based on the latest mathematical constraint parameters, a mixed-integer linear programming problem is constructed with the objective of minimizing the operating cost or maximizing the revenue of a virtual power plant. Solving this mixed-integer linear programming problem yields a control sequence in the future prediction time domain, and only the instruction for the first time step of this sequence is issued. At the next time step... The prediction error is calculated based on the received actual operation feedback data, and this error is used to correct the input for the next round of optimization.

[0013] The virtual power plant dispatching system based on regional electricity consumption activity prediction includes a data acquisition module, an electricity consumption activity type classification module, an electricity consumption prediction module, and an optimization dispatching module. The data acquisition module collects user load data, geographic information data, and activity-related data, including electric vehicle connection status and industrial production plans, for the target area. The electricity consumption activity type classification module constructs a profile of typical electricity consumption behavior characteristics based on geographic information data and user load data, and classifies users into different electricity consumption activity types. The electricity consumption prediction module uses a preset machine learning sub-model to predict the power change curve of each electricity consumption activity type in future time periods. The optimization dispatching module uses a model predictive control model to solve for the optimal control sequence for future time periods based on the electricity consumption activity prediction results and preset optimization objectives, generating dispatching instructions for distributed energy sources, energy storage systems, or controllable loads in the virtual power plant.

[0014] Preferably, the optimized scheduling module is configured to receive real-time status feedback from distributed energy sources within the virtual power plant, calculate the deviation between the predicted value and the actual value, and introduce the deviation as a feedback correction term into the scheduling optimization at the next moment.

[0015] Compared to existing technologies, the advantages of this invention are: by shifting from predicting "aggregate load" to predicting "classification activity", this invention reduces the uncertainty of prediction.

[0016] Based on high-fidelity activity profiles, the optimized dispatch model of virtual power plants can differentiate between rigid and flexible loads, thereby developing dispatch strategies superior to existing technologies. For example, it can more accurately calculate required reserve capacity, more effectively utilize energy storage, and more precisely implement demand response, thus maximizing the benefits of virtual power plants in the electricity market.

[0017] By accurately predicting the occurrence of high-risk loads (such as electric vehicle cluster charging), virtual power plants can prepare response strategies in advance, avoid grid impacts, and improve the reliability of the entire system. Attached Figure Description

[0018] Figure 1 is a flowchart of the virtual power plant scheduling method based on regional electricity consumption prediction proposed in this invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0020] Referring to Figure 1, the present invention provides a virtual power plant scheduling method based on regional electricity consumption activity prediction, including the following steps: S1, collecting user load data, geographic information data, and activity-related data including electric vehicle connection status and industrial production plans in the target area. The specific data includes: user load data: 15-minute or finer-grained user electricity consumption data from smart meters, distributed photovoltaic power generation data, and energy storage charging and discharging data.

[0021] Geographic information data: user's geographical location information (latitude and longitude coordinates) and distribution of points of interest (e.g., commercial areas, residential areas, industrial areas).

[0022] Activity-related data: Electric vehicle charging pile data: real-time connection status of charging piles, charging power, charging type (fast charging / slow charging), and historical charging behavior data (such as user arrival time and stay time).

[0023] Industrial / commercial data: Pre-planned production schedules for large industrial users (e.g., which production lines are started and when), operational schedules for commercial buildings, and holiday calendars.

[0024] Environmental data: meteorological data (temperature, humidity, light intensity), traffic flow data (used to help predict the location and time of electric vehicle charging).

[0025] S2. Based on the aforementioned geographic information data and user load data, construct a profile of electricity consumption activity types containing typical electricity consumption behavior characteristics, and classify users into different electricity consumption activity types. Specifically, this includes: preliminarily classifying power grid users within the region according to administrative divisions (such as streets, communities) or geographical coordinate ranges. Each power supply area is divided into two zones to ensure that each zone is relatively independent in terms of the distribution network topology.

[0026] Electricity consumption matrix for each region The joint matrix triecomposition algorithm is executed to decompose the user electricity consumption matrix into the product of three low-dimensional matrices: In the formula, It is a user feature matrix that captures specific attributes of users within that region; It is a time-series feature matrix that captures the electricity consumption patterns that change over time; It is a shared coefficient matrix used to constrain and discover common features across regions. This method can effectively separate region-specific spatial features from general temporal behavioral features.

[0027] In a preferred embodiment of the present invention, the user's electricity consumption data matrix is ​​not the raw data stream directly read from the electricity meter, but a two-dimensional matrix after cleaning, alignment, and structuring. The specific steps for constructing this matrix are as follows: First, raw load data is obtained from the smart meter terminal in the target area at fixed time steps (e.g., 1 hour), wherein the data items include: active power, reactive power, voltage, current, and corresponding timestamps.

[0028] Understandably, due to potential slight discrepancies or packet loss in the upload times of different meters, a standardization process is necessary. Specific standardization steps include: Time alignment: mapping all users' sampling points to a unified standard time axis. For data from non-standard times, linear interpolation is used to normalize it to the nearest standard time.

[0029] Missing value handling: For missing points on the time axis, if the missing length is less than a preset threshold (e.g., 1 hour), it is filled using Lagrange interpolation; if the missing length exceeds the threshold, the user's data for that day is removed or marked as abnormal and not included in the current matrix construction.

[0030] After the above processing, a two-dimensional user electricity consumption data matrix is ​​constructed for each power supply area. , ; where the elements in the matrix For the first The user's electricity meter is at the first Normalized load values ​​at each time step.

[0031] Understandably, in order to eliminate the impact of differences in electricity consumption levels among different users (such as the absolute difference between large factories and convenience stores) on subsequent clustering algorithms, it is necessary to normalize each row of the matrix (i.e., the electricity meter value for each user).

[0032] The preferred method in this scheme is the minimum-maximum normalization method, and the specific calculation formula is as follows: In the formula, For normalized values, For the first Load value per user, and These represent taking the maximum and minimum values, respectively.

[0033] The user electricity consumption matrix obtained through the above steps not only preserves the temporal fluctuation characteristics of users (represented by row vectors), but also implicitly contains the spatial correlation between users (represented by column vectors).

[0034] The improved fuzzy C-means algorithm optimized by Grey Wolf is used to cluster each row vector of the user electricity consumption matrix. Specifically, firstly, the initial cluster centers are determined using the Grey Wolf optimization algorithm.

[0035] Then, the fuzzy C-means algorithm is used to calculate the membership degree of each user to each cluster center.

[0036] Ultimately, users with similar feature vectors are grouped into the same electricity consumption behavior group.

[0037] S3. For each type of electricity consumption activity, use a pre-set machine learning sub-model to predict the power change curve of that type of electricity consumption activity in the future.

[0038] For the different types of electricity consumption activities identified above, different pre-trained models are used for prediction.

[0039] In a preferred embodiment, for a group of electric vehicle charging activity types: obtain their electric vehicle charging load sequence and corresponding traffic flow data.

[0040] The electric vehicle charging load sequence is decomposed into multiple intrinsic mode function components using the variational mode decomposition algorithm, thereby reducing the impact of nonlinear fluctuations in the sequence on prediction accuracy.

[0041] Each intrinsic mode function component and its corresponding traffic flow data are input into a bidirectional long short-term memory (LSTM) network. The bidirectional LSTM network, through two LSTM layers (forward and backward), can simultaneously capture past and future contextual information of the time series, effectively extracting the temporal dependency features of charging behavior.

[0042] In addition, an attention layer is introduced after the bidirectional long short-term memory network layer to automatically assign weights to the input features at different time steps (for example, the weight of traffic features during rush hour will be automatically increased), enhancing the model's sensitivity to key time nodes.

[0043] Finally, the prediction results of each intrinsic mode function component are superimposed and reconstructed to output the electric vehicle charging power change curve at future time.

[0044] In another preferred example, for a group of industrial production activity types: obtain the historical load, production schedule and holiday types of industrial production activities.

[0045] The data from the aforementioned industrial production activities are input into a base learner, wherein the base learner is selected as either a random forest model or an XGBoost model.

[0046] The output of the base learner is used as a new feature and input into a meta-learner (such as linear regression or a shallow neural network). The meta-learner learns the error patterns of the base model and performs weighted correction on the final prediction result, outputting a more deterministic industrial base load curve.

[0047] S4. Using the electricity consumption prediction results, solve for the optimal control sequence for future periods based on the preset optimization objective, and generate scheduling instructions for distributed energy, energy storage systems or controllable loads in the virtual power plant.

[0048] In this embodiment, the optimized scheduling model is a preferred model predictive control framework. Specifically: S401, at each scheduling time... Based on the latest electricity consumption forecasts, a mixed-integer linear programming problem is constructed with the objective of maximizing the revenue of a virtual power plant.

[0049] The revenue of the virtual power plant is specifically calculated as "total revenue" minus "total costs." Total revenue includes: revenue from selling electricity to the main grid, revenue from providing ancillary services (such as frequency regulation and backup), and revenue from service fees charged to internal users. Total costs include: the cost of purchasing electricity from the main grid, fuel costs of distributed generator units (such as gas turbines), charging and discharging losses of the energy storage system, and compensation costs incurred for reducing or shifting controllable loads. Furthermore, the objective function may include a penalty term to penalize the deviation between actual operating power and planned dispatch power, ensuring system stability.

[0050] Using the electricity consumption forecast results, optimization function constraints are determined to ensure that the scheduling results are physically feasible and meet user needs. The constraints specifically include: real-time power balance constraint: at each point in time during the forecast period, the sum of the power generation of all power sources inside the virtual power plant (including energy storage discharge and power purchase from the grid) is equal to the sum of the power consumption of all loads (including energy storage charging and power sales to the grid).

[0051] Rigid activity response constraint: For electrical activities identified as “rigid” (such as assembly line production), the power supplied must not be less than its predicted power consumption.

[0052] Physical limits of the equipment: The state of charge of the energy storage battery must always be within a safe range (e.g., 10%-90%), the output power of the inverter must not exceed its rated capacity, and the transmission power of the tie line must not exceed the limit, etc.

[0053] S402. Solve the mixed integer linear programming problem to obtain the optimal scheduling sequence for each time step in the future prediction time domain (e.g., the energy storage charging and discharging plan every 15 minutes in the next 4 hours), and issue the instruction to execute the first time step in the sequence.

[0054] The optimizer converts the issued instructions into specific control instructions for each unit in the virtual power plant (e.g., “Energy Storage ESS-1 discharges at 5MW power”, “The temperature control air conditioning load in Area 2 is interrupted by 10%”).

[0055] S403. After executing the above instructions (i.e., when entering the next time step), collect the actual operating status of the virtual power plant (such as the actual remaining energy storage capacity and the actual load value), and calculate the deviation between the "actual status" and the "predicted status" of the previous round. In the next round of optimization, this deviation is added to the model as a correction term for the initial conditions, thereby eliminating the cumulative effect of prediction errors.

[0056] The virtual power plant dispatching system based on regional electricity consumption activity prediction includes a data acquisition module, an electricity consumption activity type classification module, an electricity consumption prediction module, and an optimization dispatching module. The data acquisition module collects user load data, geographic information data, and activity-related data, including electric vehicle connection status and industrial production plans, for the target area. The electricity consumption activity type classification module constructs a profile of typical electricity consumption behavior characteristics based on geographic information data and user load data, and classifies users into different electricity consumption activity types. The electricity consumption prediction module uses a preset machine learning sub-model to predict the power change curve of each electricity consumption activity type in future time periods. The optimization dispatching module uses a model predictive control model to solve for the optimal control sequence for future time periods based on the electricity consumption activity prediction results and preset optimization objectives, generating dispatching instructions for distributed energy sources, energy storage systems, or controllable loads in the virtual power plant.

[0057] In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0058] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. A virtual power plant dispatching method based on regional electricity consumption activity forecasting, characterized in that, Includes the following steps: Collect user load data, geographic information data, and activity-related data, including electric vehicle connectivity status and industrial production plans, for the target area. Based on the geographic information data and user load data, a profile of electricity consumption activity types containing typical electricity consumption behavior characteristics is constructed, and users are divided into different electricity consumption activity types; for each electricity consumption activity type, a preset machine learning sub-model is used to predict the power change curve of that type of electricity consumption activity in future time periods. Using the electricity consumption prediction results, the optimal control sequence for future periods is solved based on a preset optimization objective, and scheduling instructions are generated for distributed energy, energy storage systems or controllable loads in the virtual power plant. The optimized scheduling model is a model predictive control model.

2. The virtual power plant dispatching method based on regional electricity consumption prediction according to claim 1, characterized in that, The process of constructing a profile of electricity consumption activity types containing typical electricity consumption behavior characteristics based on the geographic information data and user load data, and classifying users into different electricity consumption activity types, specifically includes: dividing the power grid into multiple topologically independent power supply areas based on the geographic information data; obtaining the user electricity consumption matrix of each power supply area, and separating the user feature matrix from the user electricity consumption matrix using a joint matrix tri-decomposition algorithm; and clustering multiple groups of electricity consumption activity types with similar electricity consumption behaviors based on the user feature matrix using a fuzzy C-means algorithm optimized and improved by Grey Wolf.

3. The virtual power plant dispatching method based on regional electricity consumption prediction according to claim 2, characterized in that, The process of obtaining the user electricity consumption matrix for each power supply area specifically includes: acquiring the raw load data of the power supply area at fixed time steps, wherein the raw load data includes active power, reactive power, voltage, current and corresponding timestamps; performing time alignment, missing value processing, outlier removal and normalization preprocessing on the raw load data; and constructing the user electricity consumption data matrix based on the preprocessed data. In the formula, Indicates the first A power consumption data matrix for each power supply area, containing elements of the matrix. For the first The user in the first Normalized load values ​​at each time step.

4. The virtual power plant dispatching method based on regional electricity consumption prediction according to claim 1, characterized in that, The machine learning sub-model includes an electric vehicle activity prediction sub-model and an industrial production activity prediction sub-model. The electric vehicle activity prediction sub-model uses variational mode decomposition combined with a bidirectional long short-term memory network to receive historical charging data and traffic flow data, and outputs an electric vehicle charging power change curve. The industrial production activity prediction sub-model uses an ensemble learning algorithm to receive production plan data and output an industrial base load curve.

5. The virtual power plant dispatching method based on regional electricity consumption prediction according to claim 1, characterized in that, The process involves solving for the optimal control sequence for future time periods based on a preset optimization objective, and generating scheduling instructions for distributed energy sources, energy storage systems, or controllable loads in the virtual power plant. Specifically, this includes: at each scheduling time... Based on the latest mathematical constraint parameters, a mixed-integer linear programming problem is constructed with the objective of minimizing the operating cost or maximizing the revenue of a virtual power plant. Solving this mixed-integer linear programming problem yields a control sequence in the future prediction time domain, and an instruction is issued to execute the first time step of this sequence. At the next time step... The prediction error is calculated based on the received actual operation feedback data, and this error is used to correct the input for the next round of optimization.

6. A virtual power plant dispatching system based on regional electricity consumption forecasting, used to implement the virtual power plant dispatching method based on regional electricity consumption forecasting as described in claims 1-5, characterized in that, The system includes a data acquisition module, an electricity consumption activity type classification module, an electricity consumption forecasting module, and an optimization scheduling module. The data acquisition module collects user load data, geographic information data, and activity-related data, including electric vehicle connection status and industrial production plans, for the target area. The electricity consumption activity type classification module constructs a profile of typical electricity consumption behavior characteristics based on geographic information data and user load data, and classifies users into different electricity consumption activity types. The electricity consumption forecasting module uses a preset machine learning sub-model to predict the power change curve of each type of electricity consumption activity in future time periods. The optimization scheduling module uses a model predictive control model to solve for the optimal control sequence for future time periods based on the electricity consumption prediction results and preset optimization objectives, generating scheduling instructions for distributed energy sources, energy storage systems, or controllable loads in the virtual power plant.

7. The virtual power plant dispatching system based on regional electricity consumption prediction according to claim 6, characterized in that, The optimized scheduling module is configured to receive real-time status feedback from distributed energy sources within the virtual power plant, calculate the deviation between the predicted and actual values, and introduce the deviation as a feedback correction term into the scheduling optimization at the next moment.

Images