Virtual power plant output prediction method and system

Abstract

The application discloses a virtual power plant output prediction method and system, and belongs to the technical field of power management. The method comprises the following steps: acquiring a standardized feature data set, which is realized by preprocessing multi-source time series operation data; constructing a light-weight time series prediction model based on a simplified gate recurrent unit and a time series attention module; inputting the data set into the model to obtain an initial output prediction result; identifying an operation condition according to a historical prediction error law; and performing hierarchical adaptive bias correction on the initial result based on the operation condition by using a differentiated correction strategy to obtain an optimal output prediction result. The application solves the problems of weak static feature extraction capability of the model, large error in mutation scene and difficulty in light-weight deployment by constructing a light-weight model and embedding an attention mechanism to strengthen key feature extraction, combining a differentiated correction strategy based on condition identification, and improving the accuracy and robustness of virtual power plant output prediction.

Description

Technical Field

[0001] This application relates to the field of power management technology, and in particular to a method and system for predicting the output of a virtual power plant. Background Technology

[0002] With the advancement of new power system construction, a large number of distributed renewable energy sources and flexible loads are aggregated and connected to the grid in the form of virtual power plants. The accuracy of their output forecasting directly affects the power balance of the power grid, the level of renewable energy absorption, and the revenue from electricity market transactions, making it a key technology for ensuring the safe and economical operation of virtual power plants. Existing technologies mostly employ time-series forecasting models driven by historical data, such as Recurrent Neural Networks (RNNs) and Long Short-Term Memory (LSTM) networks, but these have significant shortcomings in practical applications. On the one hand, traditional models are complex in structure, have a large number of parameters, and high computing power, making it difficult to meet the lightweight deployment requirements of low power consumption and low latency for virtual power plant edge terminals. On the other hand, the model feature extraction method is static and fixed, and the time series features are assigned equal weights, making it difficult to adaptively capture sudden disturbances such as weather changes, equipment start-up and shutdown, and load peak-valley switching, resulting in significant prediction lag. In addition, the error correction methods are rigid, mostly using fixed thresholds or single deviation compensation, and cannot adaptively adjust the correction strategy according to the operating conditions.

[0003] In summary, existing technologies struggle to achieve adaptive extraction and dynamic error correction of abrupt changes under complex operating conditions while ensuring lightweight model deployment. This results in low prediction accuracy and poor robustness in fluctuating scenarios, failing to meet the engineering application requirements for ultra-short-term real-time scheduling of virtual power plants. Summary of the Invention

[0004] This application provides a method and system for predicting the output of a virtual power plant, which at least solves the problems of weak static feature extraction capability, large error in sudden change scenarios, rigid correction mechanism, and difficulty in lightweight deployment of virtual power plant output prediction models in the prior art.

[0005] In a first aspect, this application provides a method for predicting the output of a virtual power plant, the method comprising: A standardized feature dataset is obtained by preprocessing the multi-source time-series operation data of the virtual power plant. The multi-source time-series operation data includes historical output data of the virtual power plant, numerical weather forecast data, power grid dispatch command data, and market electricity price data. A lightweight time-series prediction model is constructed based on a simplified gated loop unit and a time-series attention module, wherein the time-series attention module is embedded in the back end of the simplified gated loop unit. The standardized feature dataset is input into the lightweight time series prediction model to obtain the initial output prediction results of the virtual power plant. Based on historical prediction error patterns, identify the current operating conditions of the virtual power plant; Based on the operating conditions, a differentiated correction strategy is adopted to perform hierarchical adaptive deviation correction on the initial output prediction results to obtain the optimal output prediction results of the virtual power plant. The differentiated correction strategy includes: when the operating conditions are in a steady state, fixed compensation is performed using historical average deviation; when the operating conditions are under sudden disturbance, dynamic fitting correction is performed by matching similar error curves in the historical error feature library.

[0006] By employing the above technical solutions, the model complexity is reduced by using simplified gated loop units, and key feature extraction is enhanced by combining a temporal attention module, thus achieving lightweight and high-precision prediction of the model. At the same time, by identifying operating conditions and matching differentiated correction strategies, the system deviation under steady-state conditions and random fluctuation under abrupt operating conditions are effectively solved, significantly improving the accuracy and robustness of virtual power plant output prediction.

[0007] Secondly, this application provides a virtual power plant output prediction system, the system comprising: The data processing module is used to acquire a standardized feature dataset, which is achieved by preprocessing the acquired multi-source time-series operation data of the virtual power plant. The multi-source time-series operation data includes historical output data of the virtual power plant, numerical weather forecast data, power grid dispatch instruction data, and market electricity price data. The model building module is used to construct a lightweight time-series prediction model based on a simplified gated loop unit and a time-series attention module, wherein the time-series attention module is embedded in the back end of the simplified gated loop unit; the standardized feature dataset is input into the lightweight time-series prediction model to obtain the initial output prediction results of the virtual power plant; The error correction module is used to identify the current operating condition of the virtual power plant based on historical prediction error patterns. Based on the operating condition, a differentiated correction strategy is used to perform hierarchical adaptive deviation correction on the initial output prediction result to obtain the optimal output prediction result of the virtual power plant. The differentiated correction strategy includes: when the operating condition is in a steady state, fixed compensation is performed using historical average deviation; when the operating condition is under sudden disturbance, dynamic fitting correction is performed by matching similar error curves in the historical error feature library.

[0008] Thirdly, this application provides an electronic device comprising one or more processors and one or more memories, wherein at least one piece of program code is stored in the one or more memories, the program code being loaded and executed by the one or more processors to implement the operations performed by the virtual power plant output prediction method.

[0009] Fourthly, this application also provides a computer-readable storage medium storing at least one piece of program code, which is loaded and executed by a processor to implement the operations performed by the virtual power plant output prediction method.

[0010] Fifthly, this application also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of any of the above-described virtual power plant output prediction methods. Attached Figure Description

[0011] The accompanying drawings, which are incorporated in and form a part of this specification, illustrate embodiments consistent with this disclosure and, together with the description, serve to explain the principles of this disclosure.

[0012] To more clearly illustrate the technical solutions in the embodiments of this disclosure or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0013] Figure 1 A flowchart illustrating a virtual power plant output prediction method provided in this application embodiment. Figure 1 ; Figure 2 A flowchart illustrating a virtual power plant output prediction method provided in this application embodiment. Figure 2 ; Figure 3 A flowchart illustrating a virtual power plant output prediction method provided in this application embodiment. Figure 3 ; Figure 4 A flowchart illustrating a virtual power plant output prediction method provided in this application embodiment. Figure 4 ; Figure 5 A flowchart illustrating a virtual power plant output prediction method provided in this application embodiment. Figure 5 ; Figure 6 A flowchart illustrating a virtual power plant output prediction method provided in this application embodiment. Figure 6 ; Figure 7 A flowchart illustrating a virtual power plant output prediction method provided in this application embodiment. Figure 7 ; Figure 8 A schematic diagram of the structure of a virtual power plant output prediction system provided in this application embodiment; Figure 9 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0014] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the protection scope of this application.

[0015] It should be noted that, in the description of this application, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. The terms "first," "second," etc., in this application are used to distinguish similar objects and are not used to describe a specific order or sequence.

[0016] In related technologies, with the advancement of new power system construction, a large number of distributed renewable energy sources and flexible loads are aggregated and connected in the form of virtual power plants. The accuracy of their output prediction directly affects the power balance of the power grid, the level of renewable energy consumption, and the revenue from electricity market transactions. Existing technologies mostly adopt time-series prediction models driven by historical data, such as recurrent neural networks (RNN) and long short-term memory networks (LSTM). However, these models have significant drawbacks in practical applications: First, traditional models have complex structures, large numbers of parameters, and high computational costs, making it difficult to meet the lightweight deployment requirements of low power consumption and low latency for virtual power plant edge terminals. Second, the model feature extraction methods are static and fixed, and time-series features are assigned equal weights, making it difficult to adaptively capture sudden disturbances such as weather changes, equipment start-up and shutdown, and load peak-valley switching, resulting in significant prediction lag. Third, error correction methods are rigid, often using fixed thresholds or single deviation compensation, which cannot adaptively adjust correction strategies according to dynamic changes in operating conditions, resulting in low prediction accuracy and poor robustness in fluctuating scenarios, making it difficult to meet the engineering application requirements of ultra-short-term real-time dispatch of virtual power plants.

[0017] To address the aforementioned technical challenges, this application presents a virtual power plant output prediction method and system that integrates a lightweight model and adaptive error correction. First, a lightweight time-series prediction model based on simplified gated recurrent units is constructed. This reduces computational resource consumption by simplifying network structure parameters, making it suitable for lightweight edge deployment. A time-series attention module is embedded in the model's backend, dynamically calculating feature importance scores and time-series weights to enhance the extraction of key features at power surge moments, thus resolving prediction lag issues. Second, a hierarchical adaptive error correction mechanism based on operating condition identification is designed. By analyzing the fluctuation amplitude and statistical characteristics of historical prediction error sequences, steady-state and sudden disturbance operating conditions are identified, and differentiated correction strategies are applied: under steady-state conditions, fixed compensation is performed using historical average deviations; under sudden disturbance conditions, dynamic fitting correction is performed using a historical error feature library. This eliminates systemic biases and random fluctuations under different operating conditions, effectively improving the accuracy and robustness of virtual power plant output prediction.

[0018] The application scenarios of the technical solutions provided in the embodiments of this application are described below.

[0019] The technical solutions provided in this application are mainly applied to the operation and dispatching scenarios of virtual power plants that aggregate distributed energy, energy storage, and flexible loads under the background of new power systems. They aim to solve the problems of low prediction accuracy and difficulty in real-time dispatch caused by frequent interactions between power sources, grids, loads, and storage, and large fluctuations in new energy sources. Specific application scenarios are as follows: 1. Real-time edge prediction scenario for industrial park-level virtual power plants In industrial park microgrids that include distributed photovoltaic power, decentralized wind power, user-side energy storage, charging piles, and industrial flexible loads, edge computing terminals (such as gateways) often have limited computing resources. Traditional deep learning models have a large number of parameters and slow inference, making it difficult to run them in real time at the edge.

[0020] This application employs a simplified gated cyclic unit to construct a lightweight time-series prediction backbone model. By removing redundant structures such as batch normalization and random deactivation, the hidden layer dimension is compressed to 64 dimensions, reducing computational overhead. This technical solution can be directly deployed on the edge gateway of an industrial park, achieving millisecond-level local inference and meeting the timeliness requirements of industrial parks for minute-level real-time monitoring and on-site consumption of new energy output.

[0021] 2. Supporting Scenarios for Electricity Spot Market Trading Auctions In the electricity spot market, virtual power plants need to submit bids and applications based on predicted output. The accuracy of the prediction directly determines market revenue and default costs (deviation losses). When faced with sudden weather changes, such as cloud cover or sudden changes in wind speed, traditional model predictions lag significantly, easily leading to huge deviation assessment fees.

[0022] The hierarchical adaptive deviation correction strategy designed in this application can automatically identify whether the current operating condition is in a steady state or under sudden disturbance based on historical error patterns. During periods of stable weather, fixed compensation is used to eliminate system deviations, while during periods of extreme weather changes, dynamic fitting and correction are performed using a historical error feature library. In this scenario, this application can significantly improve the accuracy of ultra-short-term forecasts, helping virtual power plants accurately declare electricity volume in the spot market, reduce deviation settlement costs, and maximize market transaction revenue.

[0023] 3. Regional-level virtual power plant aggregator dispatching auxiliary decision-making scenario For virtual power plant aggregators that aggregate a large number of scattered distributed resources, accurate power output plans need to be reported to the power grid dispatch center. However, the output of new energy sources often fluctuates drastically due to sudden weather changes. This application, by embedding a time-series attention module, can adaptively identify and strengthen the feature weights of critical moments such as sudden weather changes, load peak-valley switching, and equipment start-up and shutdown.

[0024] In dispatching and decision-making scenarios, this technical solution can effectively capture the trend of power surges or drops, assist dispatchers in predicting power fluctuations in advance, formulate more reasonable energy storage charging and discharging strategies and load shifting schemes, avoid dispatching deviation assessments caused by prediction lags, and improve the grid's ability to accept new energy sources.

[0025] 4. Source-grid-load-storage coordinated control and emergency response scenarios When the power grid issues dispatch instructions or frequency anomalies require virtual power plants to participate in ancillary services, the system needs to respond quickly. This application, through multi-source data fusion including power grid dispatch instruction data, can respond to power grid dispatch needs in a timely manner.

[0026] By combining high-precision optimal output prediction results, the system can accurately calculate the state of charge / discharge adjustment of the energy storage system and the power allocation ratio of distributed energy. For example, during periods of high renewable energy generation, the system prioritizes full consumption of surplus electricity and stores it in energy storage; during periods of low output, it releases stored electricity to make up for the shortfall. In this scenario, the technical solution of this application ensures the accuracy and response speed of the virtual power plant when participating in grid peak shaving and frequency regulation ancillary services.

[0027] In summary, the technical solution of this application covers the entire process from lightweight deployment at the edge and feature extraction in high-fluctuation scenarios to adaptive error correction under multiple operating conditions, which can effectively support the accurate prediction and intelligent scheduling of virtual power plants in complex operating environments.

[0028] After introducing the implementation environment of the embodiments of this application, the technical solutions provided by the embodiments of this application are described below. (See also...) Figure 1 Taking a virtual power plant dispatching system as the executing entity as an example, the method includes the following steps: Step S101: Obtain a standardized feature dataset by preprocessing the acquired multi-source time-series operation data of the virtual power plant.

[0029] The multi-source time-series operational data includes historical output data of virtual power plants, numerical weather forecast data, power grid dispatch command data, and market electricity price data.

[0030] Data is the foundation of predictive models, and its quality directly affects prediction accuracy. The multi-source time-series operational data originates from distributed photovoltaic, wind power, energy storage, and flexible load resources aggregated by the virtual power plant, and is characterized by being multi-source, heterogeneous, and noisy. This step aims to generate a high-quality, standardized model input dataset through systematic preprocessing.

[0031] Specifically, the preprocessing includes the following sub-steps: First, acquire multi-source time-series operational data of the virtual power plant. Second, based on the multi-source time-series operational data of the virtual power plant, use the 3σ criterion to identify and remove outlier data. This method calculates the mean μ and standard deviation σ of the time-series data, identifies data points exceeding the range [μ-3σ, μ+3σ] as outliers caused by equipment failure, communication jitter, etc., and removes them, obtaining time-series data after anomaly removal. Third, based on the time-series data after anomaly removal, a differentiated completion strategy is adopted for missing data: for short-term single-point missing data, linear interpolation is used for smooth completion to ensure sequence continuity; for long-term continuous missing data, historical averages of the same season, day, and time period are used for completion to ensure the rationality of data trends, thereby obtaining time-series data after missing data repair. Finally, based on the time-series data after missing data repair, normalization and time-series alignment are performed. Normalization maps data of different scales to the [0,1] interval, eliminating the interference of numerical scale differences on model training; temporal alignment unifies data with different sampling frequencies and timestamps to the same time base, such as a 5-minute interval, and finally obtains a standardized feature dataset with regular temporal order and clean data, providing reliable input for subsequent model prediction.

[0032] Step S102: Construct a lightweight time series prediction model based on a simplified gated loop unit and a time series attention module.

[0033] The timing attention module is embedded in the back end of the simplified gated loop unit; Traditional recurrent neural network (RNN) models have a large number of parameters and high computational cost, making them difficult to deploy at edge environments. This step aims to resolve the contradiction between model complexity and limited deployment resources by constructing a lightweight model, while introducing an attention mechanism to enhance the ability to extract key features.

[0034] Specifically, the construction process is as follows: First, based on the simplified gated recurrent unit (GRU) and the temporal attention module, the lightweight network structure parameters are determined. The simplified GRU is implemented on the standard GRU structure by reducing the hidden layer dimension (e.g., fixing it to 64 dimensions), removing redundant structures such as batch normalization layers and random deactivation layers, thereby significantly compressing the number of model parameters and computational cost. Second, based on the lightweight network structure parameters, the hidden layer dimension and stacking layer number of the simplified GRU are determined, for example, a two-layer stacking structure is adopted to further optimize model efficiency while ensuring feature extraction capabilities. Subsequently, based on the hidden layer dimension of the simplified GRU, the feature weight allocation method of the temporal attention module is determined. This module calculates the dynamic weights at each time step of the temporal hidden feature sequence output by the GRU through operations such as fully connected transformation, tanh activation, and softmax normalization. Finally, based on the feature weight allocation method and the simplified GRU, a lightweight temporal prediction model is constructed. The temporal attention module is embedded in the back end of the simplified GRU unit, which can dynamically adjust the contribution of features at different times, enhance the extraction of key moment features such as meteorological changes and load fluctuations, and make up for the shortcomings of the equal distribution of feature weights in traditional models.

[0035] Step S103: Input the standardized feature dataset into the lightweight time series prediction model to obtain the initial output prediction result of the virtual power plant.

[0036] This step is the core of the model's feature learning and inference prediction. The specific process is as follows: The standardized feature dataset obtained in step S101 is input into the constructed lightweight time-series prediction model. The model first uses an embedded time-series attention module to determine the time-series hidden feature sequence based on the input data and calculates the feature importance score for each time-series node. Based on the feature importance scores, dynamic time-series weights are obtained through softmax normalization, assigning high weights to key time steps and low weights to stationary time steps. Finally, the time-series hidden feature sequence is weighted and fused based on the dynamic time-series weights, and then passed through a fully connected output layer to obtain the initial power output prediction result for the virtual power plant in the future period (e.g., the next 15 minutes to 2 hours). This initial power output prediction result is uncorrected and reflects the model's preliminary judgment on the power output trend.

[0037] Step S104: Identify the current operating conditions of the virtual power plant based on historical prediction error patterns.

[0038] The output of a virtual power plant is affected by various factors such as weather and load, resulting in complex and variable operating conditions. This step aims to intelligently distinguish between two typical operating conditions—"steady state" and "abrupt change"—by quantitatively analyzing historical prediction errors, providing a basis for subsequent differentiated corrections.

[0039] Specifically, firstly, based on historical prediction error patterns, a historical prediction error sequence is determined. This sequence consists of the difference between the initial historical prediction value and the actual output value. Secondly, based on the historical prediction error sequence, the error fluctuation amplitude is calculated, for example, using the standard deviation or range of the error as a metric. Subsequently, based on the error fluctuation amplitude, a fluctuation comparison result is determined. To achieve adaptive identification, this step also includes dynamically setting a preset threshold: based on the error fluctuation amplitude, statistical characteristic values ​​of the error distribution, such as mean and variance, are further calculated. Based on these statistical characteristic values, boundary thresholds for dividing operating conditions are determined, thereby obtaining a preset threshold for identifying the current operating condition. Finally, based on the fluctuation comparison result and the preset threshold, the current operating condition of the virtual power plant is identified: if the current error fluctuation amplitude is less than the threshold, it is determined to be a steady-state operating condition; otherwise, it is determined to be a sudden disturbance operating condition.

[0040] Step S105: Based on the operating conditions, a differentiated correction strategy is used to perform hierarchical adaptive deviation correction on the initial output prediction results to obtain the optimal output prediction results of the virtual power plant.

[0041] The differentiated correction strategy includes: when the operating condition is in a steady state, using the historical average deviation for fixed compensation; when the operating condition is under sudden disturbance, matching the same type of error curve in the historical error feature library for dynamic fitting correction.

[0042] Traditional, single, fixed compensation strategies are ill-suited to complex and ever-changing working conditions. The core of this step lies in dynamically selecting the most suitable correction strategy based on the working condition identification results to achieve precise deviation correction.

[0043] The specific implementation method is as follows: First, based on the identification result of step S104, the current operating condition type is determined. If the operating condition type is a steady-state operating condition, the average prediction deviation of the historical steady-state period is calculated and used as a fixed compensation amount, which is superimposed on the initial output prediction result to obtain a steady-state corrected prediction result, thereby eliminating systematic deviations and ensuring a smooth prediction curve. If the operating condition type is a sudden disturbance operating condition, the most similar historical error curve of the same type (such as fluctuation shape and amplitude) is matched from the pre-constructed historical error feature library. Then, the matched historical error curve is used as a correction template to dynamically fit and correct the initial output prediction result, generating a dynamically corrected prediction result, thereby specifically offsetting the prediction deviation caused by sudden disturbances. Finally, the corrected, high-precision, and highly robust optimal output prediction result is output.

[0044] This embodiment constructs a lightweight time-series prediction model based on a simplified gated loop unit and embeds a time-series attention module. This significantly reduces the model's computational overhead to adapt to edge deployments. At the same time, it utilizes an attention mechanism to dynamically allocate feature weights to enhance the extraction of key features at abrupt changes, thus solving the problem of significant lag in traditional models. Furthermore, by identifying differentiated correction strategies that match steady-state fixed compensation with abrupt dynamic fitting correction, it effectively eliminates system biases and random fluctuations under different operating conditions, significantly improving the accuracy and robustness of virtual power plant output prediction.

[0045] It should be noted that the above steps S101-S105 are a simplified description of the embodiments provided in this application.

[0046] The methods provided in the embodiments of this application will be described in more detail below with reference to some examples. See also Figure 2 The acquisition of the standardized feature dataset is achieved by preprocessing the acquired multi-source time-series operation data of the virtual power plant, specifically including the following steps: Step S201: Obtain multi-source time-series operation data of the virtual power plant.

[0047] Multi-source time-series operational data forms the foundation for predictive models, and its coverage dimensions determine the comprehensiveness of the model's understanding of operating conditions. This step aims to construct a raw database encompassing multi-dimensional characteristics of the power source, grid, load, and storage systems.

[0048] Specifically, with a sampling period of 5 minutes, real-time power data of distributed photovoltaic, wind power, energy storage and flexible loads aggregated by the virtual power plant are collected simultaneously, numerical weather forecast data (NWP) transmitted by the meteorological station, dispatch command data issued by the power grid and real-time electricity price data of the electricity market are collected simultaneously.

[0049] Step S202: Based on the multi-source time-series operation data of the virtual power plant, the 3σ criterion is used to identify and remove outlier time-series data to obtain time-series data after removing outliers.

[0050] Field-collected data often contains outliers due to sensor malfunctions or communication interference. Failure to remove these outliers will hinder the convergence direction of the training model. This step aims to purify the dataset using statistical methods.

[0051] Specifically, for the collected raw data, its mean μ and standard deviation σ are calculated. Data points whose values ​​fall outside the interval [μ-3σ, μ+3σ] are identified as outliers caused by equipment failure or communication interference and are removed to purify the dataset.

[0052] Step S203: Based on the time series data after removing anomalies, use linear interpolation or the historical average of the same day and time to complete the missing data, and obtain the time series data after repair.

[0053] The continuity of time-series data is a prerequisite for the model to extract time-dependent features. Missing values ​​will cause the sequence to break and affect the prediction accuracy. This step aims to fill in the data gaps and ensure the integrity of the time series.

[0054] Specifically, for short-term single-point missing data, a linear interpolation algorithm is used to smoothly complete the missing data; for continuous long-term missing data, the average normal operating data of the same season, day and time period in the historical database is retrieved to complete the missing data, ensuring the continuity and integrity of the time series.

[0055] Step S204: Based on the repaired missing time series data, perform normalization processing and time series alignment to obtain a standardized feature dataset.

[0056] Data with different scales and sampling frequencies cannot be directly computed in parallel; standardization is a necessary step for efficient model training. This step aims to eliminate data scale differences and unify the time reference.

[0057] Specifically, the Min-Max normalization method is used to map data of different dimensions to the [0,1] interval to eliminate the difference in dimensions; at the same time, data with different sampling frequencies are uniformly aligned to the same time base to form a standardized multi-source feature dataset.

[0058] This embodiment achieves refined cleaning of multi-source heterogeneous data by employing the 3σ criterion to remove outliers, a differential strategy to fill in missing values, and normalization and time-series alignment techniques. Its beneficial effects are that it effectively removes interference noise, eliminates dimensional differences, and constructs a high-quality, time-ordered standardized dataset, laying a solid data foundation for the accurate prediction of subsequent models.

[0059] In some embodiments, see Figure 3 The lightweight time series prediction model, based on a simplified gated loop unit and a temporal attention module, specifically includes the following steps: Step S301: Determine the lightweight network structure parameters based on the simplified gated recurrent unit and the temporal attention module.

[0060] The model's structural parameters directly determine its computational overhead and inference speed, and must be adapted to edge hardware resources. This step aims to establish the overall parameter framework for model construction.

[0061] Specifically, based on the hardware resource configuration of the smart terminal at the edge of the virtual power plant, such as memory size and CPU computing power, the initial parameters that the model as a whole needs to meet, such as the parameter scale and network type, are determined.

[0062] Step S302: Based on the lightweight network structure parameters, determine the constraints on the model's computational resource consumption.

[0063] Edge computing resources are limited, and specific resource constraints must be quantified to guide network design. This step aims to define hard boundaries for lightweight model design.

[0064] Specifically, based on the limitations of hardware resources, specific indicators that the quantification model inference must meet are set, such as setting the single inference latency of the model to be less than 10ms and the number of model parameters to be less than 1MB, which serve as hard constraints for subsequent structural design.

[0065] Step S303: Calculate the resource consumption constraints based on the model and determine the hidden layer dimension of the simplified gated loop unit.

[0066] Hidden layer dimension is a direct trade-off between the model's feature representation capability and computational cost. This step aims to find the optimal feature dimension under resource constraints.

[0067] Specifically, under the premise of satisfying inference latency and storage constraints, the maximum feature dimension that the backpropagation network can handle is determined. For example, under the constraint of completing inference within 10ms, experimental calculations determine that the hidden layer dimension should be set to 64 dimensions to maximize feature representation capability with limited computing power.

[0068] Step S304: Determine the number of stacking layers of the simplified gated loop unit based on the hidden layer dimension of the simplified gated loop unit.

[0069] Model depth determines the level of abstraction in feature extraction and needs to be optimized in conjunction with feature dimensions. This step aims to balance model depth and computational efficiency.

[0070] Specifically, based on the determined hidden layer dimensions, the model's depth and performance are further balanced. If the dimension is small (e.g., 64 dimensions), the number of stacking layers can be appropriately increased (e.g., set to 2 layers) to enhance the fitting ability to complex temporal patterns; if the dimension is large, the number of layers needs to be reduced to prevent overfitting and computation timeouts.

[0071] Step S305: Based on the hidden layer dimension of the simplified gated loop unit, determine the feature weight allocation method of the temporal attention module.

[0072] The core of the attention mechanism lies in the dynamic allocation of weights, and its mapping logic needs to be adapted to the dimension of the input features. This step aims to establish a normalized mapping from feature scores to weights.

[0073] Specifically, based on the dimension of the GRU output features (i.e., 64 dimensions), the number of nodes and activation function (such as softmax) of the fully connected layer inside the attention module are designed, and the mapping logic from input features to dynamic weight coefficients is established.

[0074] Step S306: Based on the feature weight allocation method and the simplified gated loop unit, construct a lightweight time series prediction model.

[0075] Building a lightweight time series forecasting model is a process of organically integrating its components. This step aims to complete the construction of the model's complete topology from input to output.

[0076] Specifically, the designed temporal attention module is embedded in the back end of the simplified gated loop unit, connected to the output layer, and combined to form the final lightweight prediction model architecture.

[0077] This embodiment achieves a lightweight design of the model structure by reverse-determining the dimension and number of hidden layers of the model based on resource constraints and embedding an attention module. Its beneficial effect is that it significantly reduces the computational complexity and storage requirements of the model, enabling it to be efficiently deployed on edge devices with limited computing power, while retaining the ability to enhance key feature extraction.

[0078] In some embodiments, see Figure 4 The step of inputting the standardized feature dataset into the lightweight time-series prediction model to obtain the initial output prediction result of the virtual power plant specifically includes the following steps: Step S401: Input the standardized feature dataset into the lightweight time series prediction model to determine the time series hidden feature sequence.

[0079] The hidden feature sequence is a deep encoding of the original input data, containing state information from historical moments. This step aims to extract feature vectors with temporal dependencies.

[0080] Specifically, the standardized feature dataset flows through a simplified GRU layer to extract the hidden state sequence containing temporal dependencies.

[0081] Step S402: Determine the feature importance score based on the temporal hidden feature sequence.

[0082] The contribution of features to the prediction results varies at different times, and their importance needs to be quantified. This step aims to calculate the raw contribution value of features at each time point using an attention network.

[0083] Specifically, the raw feature importance score for each time node is calculated through the fully connected transformation and tanh activation function within the attention module.

[0084] Step S403: Determine the dynamic temporal weights based on the feature importance scores.

[0085] To achieve adaptive feature selection, the scores need to be converted into weights in the form of a probability distribution. This step aims to generate a normalized attention weight vector.

[0086] Specifically, the original scores are normalized using the softmax function to generate a dynamic time-series weight vector with a sum of 1, which assigns high weights to critical moments such as sudden weather changes and load fluctuations.

[0087] Step S404: Based on the dynamic time series weights and the time series hidden feature sequence, obtain the initial output prediction result of the virtual power plant.

[0088] Weighted fusion is a key step in achieving enhancement of critical features and suppression of stable features. This step aims to output the model's preliminary judgment on the future.

[0089] Specifically, the weight vector and the hidden feature sequence are weighted and summed, and then mapped through the output layer to obtain the initial power output prediction value for future periods. This embodiment achieves adaptive selection of input features by calculating feature importance scores and generating dynamic temporal weights for weighted prediction. Its beneficial effect is that it can adaptively focus on key time nodes that affect power fluctuations, improving the model's ability to capture abrupt changes and its prediction accuracy.

[0090] In some embodiments, see Figure 5 The process of identifying the current operating condition of the virtual power plant based on historical prediction error patterns specifically includes the following steps: Step S501: Determine the historical prediction error sequence based on the historical prediction error pattern.

[0091] Historical error sequences reflect the model's performance characteristics under different operating conditions and serve as a data source for operating condition identification. This step aims to construct a historical sample library for analysis.

[0092] Specifically, a rolling historical error database is maintained to record the deviation between the initial predicted value and the actual value in past periods.

[0093] Step S502: Determine the error fluctuation amplitude based on the historical prediction error sequence.

[0094] Volatility is a direct indicator of operational stability. This step aims to quantify the degree of error dispersion over historical periods.

[0095] Specifically, the standard deviation or range of the error sequence within a preset time window is calculated as an indicator to measure the magnitude of fluctuation.

[0096] Step S503: Based on the error fluctuation amplitude, determine the statistical characteristic value of the error distribution.

[0097] A single fluctuation index is insufficient to fully describe the characteristics of the operating conditions; therefore, statistical features must be introduced for in-depth analysis. This step aims to uncover the underlying patterns in the error distribution.

[0098] Specifically, the statistical characteristics of the error distribution, such as skewness and kurtosis, are further calculated to determine the shape of the error distribution.

[0099] Step S504: Determine the boundary threshold for dividing the working conditions based on the statistical characteristic values ​​of the error distribution.

[0100] Fixed thresholds are difficult to adapt to changing operating environments, while dynamic thresholds are more robust. This step aims to adaptively set the decision boundary based on statistical data.

[0101] Specifically, the boundary thresholds for distinguishing between steady state and sudden change are dynamically adjusted based on statistical characteristic values, rather than using fixed values, in order to adapt to different seasons and operating stages.

[0102] Step S505: Based on the boundary thresholds of the operating condition division, set the judgment rules for steady-state operating conditions and sudden disturbance operating conditions.

[0103] The judgment rules are the logical core for achieving automatic classification of working conditions. This step aims to establish clear classification logic criteria.

[0104] Specifically, the following rules are set: if the current error fluctuation amplitude is less than the boundary threshold, it is determined to be a steady-state condition; otherwise, it is determined to be a sudden disturbance condition.

[0105] This embodiment achieves intelligent adaptive identification of operating conditions by dynamically determining the threshold for classifying operating conditions through mining the statistical feature values ​​of historical errors. Its beneficial effect is that it avoids the problem of poor adaptability of fixed thresholds under different operating conditions, and can accurately distinguish between stable operation and drastic fluctuation scenarios, providing an accurate decision-making basis for subsequent differentiated correction.

[0106] In some embodiments, see Figure 6 The step of applying a differentiated correction strategy to perform hierarchical adaptive deviation correction on the initial output prediction results based on the operating conditions to obtain the optimal output prediction results of the virtual power plant specifically includes the following steps: Step S601: Determine the current operating condition type based on the operating conditions.

[0107] The operating conditions include steady-state operating conditions and sudden disturbance operating conditions.

[0108] Error characteristics vary greatly under different operating conditions and must be handled in a categorized manner. This step aims to clarify the type of operating scenario currently in which we are operating.

[0109] Specifically, based on the identification results, it is determined whether the current operating condition is a steady state or a sudden disturbance condition.

[0110] Step S602: When the operating condition type is a steady-state operating condition, determine the historical average deviation based on historical prediction error data.

[0111] Under steady-state conditions, system deviation dominates, and the average deviation can effectively reflect the systematic offset. This step aims to calculate the reference correction under steady-state conditions.

[0112] Specifically, error data is retrieved from the steady-state historical database, and its average value is calculated as the systematic deviation.

[0113] Step S603: Based on the historical average deviation, perform fixed compensation on the initial output prediction result to obtain the steady-state corrected prediction result.

[0114] Fixed compensation strategies are simple and efficient, and suitable for handling systematic biases in stationary sequences. This step aims to eliminate constant errors under steady-state conditions.

[0115] Specifically, the historical average deviation is superimposed on the initial forecast value to eliminate systematic errors and obtain a smooth correction result.

[0116] Step S604: When the operating condition type is a sudden disturbance operating condition, match the same type of error curve based on the historical error feature library.

[0117] The errors caused by sudden changes in operating conditions are highly random, requiring targeted corrections by identifying similar historical scenarios. This step aims to utilize a historical experience database for similarity matching.

[0118] Specifically, the historical error curve with the highest similarity to the current error waveform is retrieved from the mutation error feature library.

[0119] Step S605: Based on the aforementioned similar error curves, dynamically fit and correct the initial output prediction result to obtain a dynamically corrected prediction result.

[0120] Dynamic fitting can adapt to the nonlinear changes in abrupt error. This step aims to use the matching curve for precise correction.

[0121] Specifically, by utilizing the historical curve evolution patterns obtained through matching, dynamic correction values ​​are generated to make targeted corrections to the initial predicted values.

[0122] Step S606: Based on the steady-state correction prediction result or the dynamic correction prediction result, obtain the optimal output prediction result of the virtual power plant.

[0123] The optimal output is the endpoint of the prediction process. This step aims to summarize and output the corrected results.

[0124] Specifically, the output is the final predicted value after differentiation processing.

[0125] This embodiment achieves hierarchical adaptive error correction by distinguishing between steady-state and abrupt change conditions and employing fixed compensation and dynamic fitting correction techniques respectively. Its beneficial effect is that it implements the most suitable correction strategy for different operating conditions, which not only ensures smoothness under steady-state conditions but also solves the problem of single correction failure under abrupt change scenarios, thus comprehensively improving the accuracy and robustness of the prediction results.

[0126] In some embodiments, see Figure 7 After obtaining the optimal output prediction result of the virtual power plant, the method further includes: Step S701: Based on the optimal output prediction results, generate dispatch control instructions for the virtual power plant.

[0127] The prediction results need to be translated into specific instructions to guide the equipment's actions. This step aims to convert numerical predictions into machine instructions.

[0128] Specifically, the corrected high-precision predictions are converted into executable scheduling and control instructions.

[0129] Step S702: Determine the power allocation ratio of distributed energy sources based on the scheduling control instructions.

[0130] Power allocation is the core of achieving multi-energy collaborative optimization. This step aims to calculate the output share of each distributed unit.

[0131] Specifically, based on the predicted power gap or surplus, the output ratio that each unit, such as photovoltaic, wind power, and energy storage, should undertake is calculated.

[0132] Step S703: Adjust the charging and discharging state of the energy storage system based on the power allocation ratio of the distributed energy source.

[0133] Energy storage regulation is an important means of balancing the fluctuations in new energy sources. This step aims to implement specific charge and discharge control strategies.

[0134] Specifically, the energy storage converter is controlled according to the allocation ratio, charging and consuming during periods of high renewable energy generation, and discharging to support energy during off-peak periods, thereby achieving efficient energy utilization.

[0135] This embodiment transforms forecast data into scheduling actions, which effectively improves the economic operation of virtual power plants and ensures power balance and renewable energy absorption capacity of the power grid.

[0136] Figure 8 This is a schematic diagram of the structure of a virtual power plant output prediction system provided in an embodiment of this application. See also... Figure 8 The virtual power plant output prediction system 8 includes: a data processing module 801, a model building module 802, and an error correction module 803.

[0137] The data processing module 801 is used to acquire a standardized feature dataset, which is achieved by preprocessing the acquired multi-source time-series operation data of the virtual power plant. The multi-source time-series operation data includes historical output data of the virtual power plant, numerical weather forecast data, power grid dispatch instruction data, and market electricity price data. The model building module 802 is used to build a lightweight time-series prediction model based on a simplified gated loop unit and a time-series attention module, wherein the time-series attention module is embedded in the back end of the simplified gated loop unit; the standardized feature dataset is input into the lightweight time-series prediction model to obtain the initial output prediction result of the virtual power plant; The error correction module 803 is used to identify the current operating condition of the virtual power plant based on historical prediction error patterns; based on the operating condition, a differentiated correction strategy is used to perform hierarchical adaptive deviation correction on the initial output prediction result to obtain the optimal output prediction result of the virtual power plant. The differentiated correction strategy includes: when the operating condition is in a steady state, fixed compensation is performed using historical average deviation; when the operating condition is under sudden disturbance, dynamic fitting correction is performed by matching similar error curves in the historical error feature library.

[0138] To translate forecast results into actual control strategies, refer to... Figure 8 The virtual power plant output prediction system 8 also includes a scheduling output module 804.

[0139] The scheduling output module 804 is used to generate scheduling control instructions for the virtual power plant based on the optimal output prediction results; determine the power allocation ratio of the distributed energy source based on the scheduling control instructions; and adjust the charging and discharging state of the energy storage system based on the power allocation ratio of the distributed energy source.

[0140] It should be noted that the virtual power plant output prediction system provided in the above embodiments is only illustrated by the division of the above functional modules. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the computer equipment can be divided into different functional modules to complete all or part of the functions described above. In addition, the virtual power plant output prediction system and the virtual power plant output prediction method embodiments provided in the above embodiments belong to the same concept, and the specific implementation process is detailed in the method embodiments, which will not be repeated here.

[0141] This application also provides an electronic device. Figure 9 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application.

[0142] Typically, an electronic device 900 includes one or more processors 901 and one or more memories 902.

[0143] Processor 901 may include one or more processing cores, such as a quad-core processor, a hexa-core processor, etc. Processor 901 may be implemented using at least one hardware form selected from DSP (Digital Signal Processing), FPGA (Field-Programmable Gate Array), and PLA (Programmable Logic Array). Processor 901 may also include a main processor and a coprocessor. The main processor, also known as a CPU (Central Processing Unit), is used to process data in the wake-up state; the coprocessor is a low-power processor used to process data in the standby state. In some embodiments, processor 901 may integrate a GPU (Graphics Processing Unit), which is responsible for rendering and drawing the content to be displayed on the screen. In some embodiments, processor 901 may also include an AI (Artificial Intelligence) processor, which is used to handle computational operations related to machine learning and can be used to perform inference operations of the lightweight time-series prediction model in this application.

[0144] The memory 902 may include one or more computer-readable storage media, which may be non-transitory. The memory 902 may also include high-speed random access memory and non-volatile memory, such as one or more disk storage devices or flash memory devices. In some embodiments, the non-transitory computer-readable storage media in the memory 902 are used to store at least one computer program, which is executed by the processor 901 to implement the virtual power plant output prediction method provided in the method embodiments of this application.

[0145] Those skilled in the art will understand that Figure 9 The structure shown does not constitute a limitation on the electronic device 900, and may include more or fewer components than shown, or combine certain components, or use different component arrangements.

[0146] In addition, the device provided in the embodiments of this application may specifically be a chip, component or module. The chip may include a connected processor and a memory. The memory is used to store instructions. When the processor calls and executes the instructions, the chip can execute a virtual power plant output prediction method provided in the above embodiments.

[0147] This embodiment also provides a computer-readable storage medium storing computer program code. When the computer program code is run on a computer, the computer executes the above-described related method steps to implement the virtual power plant output prediction method provided in the above embodiment.

[0148] This embodiment also provides a computer program product that, when run on a computer, causes the computer to perform the aforementioned related steps to implement the virtual power plant output prediction method provided in the above embodiment.

[0149] In this embodiment, the device, computer-readable storage medium, computer program product, or chip are all used to execute the corresponding methods provided above. Therefore, the beneficial effects they can achieve can be referred to the beneficial effects in the corresponding methods provided above, and will not be repeated here.

[0150] Through the above description of the embodiments, those skilled in the art will understand that, for the sake of convenience and brevity, only the division of the above functional modules is used as an example. In actual applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above.

[0151] In the embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of modules or units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another apparatus, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0152] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for predicting the output of a virtual power plant, characterized in that, The method includes: A standardized feature dataset is obtained by preprocessing the multi-source time-series operation data of the virtual power plant. The multi-source time-series operation data includes historical output data of the virtual power plant, numerical weather forecast data, power grid dispatch command data, and market electricity price data. A lightweight time-series prediction model is constructed based on a simplified gated loop unit and a time-series attention module, wherein the time-series attention module is embedded in the back end of the simplified gated loop unit. The standardized feature dataset is input into the lightweight time series prediction model to obtain the initial output prediction results of the virtual power plant. Based on historical prediction error patterns, identify the current operating conditions of the virtual power plant; Based on the operating conditions, a differentiated correction strategy is adopted to perform hierarchical adaptive deviation correction on the initial output prediction results to obtain the optimal output prediction results of the virtual power plant. The differentiated correction strategy includes: when the operating conditions are in a steady state, fixed compensation is performed using historical average deviation; when the operating conditions are under sudden disturbance, dynamic fitting correction is performed by matching similar error curves in the historical error feature library.

2. The method according to claim 1, characterized in that, The acquisition of the standardized feature dataset is achieved by preprocessing the acquired multi-source time-series operational data of the virtual power plant, including: Acquire multi-source time-series operation data of a virtual power plant; Based on the multi-source time-series operation data of the virtual power plant, the 3σ criterion is used to identify and remove outlier time-series data to obtain time-series data after removing outliers. Based on the time series data after removing anomalies, the missing data is filled in using linear interpolation or the historical average of the same day and time to obtain the time series data after repair. Based on the repaired missing time series data, normalization and time series alignment are performed to obtain a standardized feature dataset.

3. The method according to claim 1, characterized in that, The lightweight time series prediction model, based on a simplified gated loop unit and a temporal attention module, includes: Based on a simplified gated recurrent unit and a temporal attention module, the lightweight network structure parameters are determined. Based on the lightweight network structure parameters, the hidden layer dimension and stacking number of simplified gated loop units are determined; Based on the hidden layer dimension of the simplified gated loop unit, the feature weight allocation method of the temporal attention module is determined; A lightweight time series prediction model is constructed based on the aforementioned feature weight allocation method and the simplified gated loop unit.

4. The method according to claim 3, characterized in that, The step of determining the hidden layer dimension and stacking number of simplified gated recurrent units based on the lightweight network structure parameters includes: Based on the lightweight network structure parameters, the constraints on the model's computational resource consumption are determined. Based on the constraints of resource consumption calculated by the model, the hidden layer dimension of the simplified gated loop unit is determined; The number of stacking layers of the simplified gated loop unit is determined based on the hidden layer dimension of the simplified gated loop unit.

5. The method according to claim 1, characterized in that, The step of inputting the standardized feature dataset into the lightweight time-series prediction model to obtain the initial output prediction result of the virtual power plant includes: The standardized feature dataset is input into the lightweight time series prediction model to determine the time series hidden feature sequence; Based on the temporal hidden feature sequence, the feature importance score is determined; Based on the feature importance scores, determine the dynamic temporal weights; Based on the dynamic temporal weights and the temporal hidden feature sequence, the initial output prediction result of the virtual power plant is obtained.

6. The method according to claim 1, characterized in that, The process of identifying the current operating conditions of the virtual power plant based on historical prediction error patterns includes: Based on the historical prediction error patterns, determine the historical prediction error sequence; Based on the historical prediction error sequence, the error fluctuation amplitude is determined; Based on the error fluctuation amplitude, the fluctuation comparison result is determined; Based on the fluctuation comparison results and the preset threshold, the current operating conditions of the virtual power plant are identified.

7. The method according to claim 6, characterized in that, After determining the error fluctuation amplitude based on the historical prediction error sequence, the method further includes: Based on the error fluctuation amplitude, the statistical characteristic values ​​of the error distribution are determined; Based on the statistical characteristic values ​​of the error distribution, the boundary threshold for dividing the working conditions is determined; Based on the boundary thresholds for the aforementioned operating conditions, rules for determining steady-state operating conditions and sudden disturbance operating conditions are set.

8. The method according to claim 1, characterized in that, Based on the operating conditions, a differentiated correction strategy is used to perform hierarchical adaptive deviation correction on the initial output prediction results to obtain the optimal output prediction results of the virtual power plant, including: Based on the operating conditions, the current operating condition type is determined, which includes steady-state operating conditions and sudden disturbance operating conditions; When the operating condition is a steady-state condition, the historical average deviation is determined based on historical prediction error data; Based on the historical average deviation, the initial output prediction result is compensated for by a fixed method to obtain the steady-state corrected prediction result. When the operating condition type is a sudden disturbance operating condition, the same type of error curve is matched based on the historical error feature library; Based on the aforementioned similar error curves, the initial output prediction result is dynamically fitted and corrected to obtain a dynamically corrected prediction result. Based on the steady-state correction prediction result or the dynamic correction prediction result, the optimal output prediction result of the virtual power plant is obtained.

9. The method according to claim 1, characterized in that, After obtaining the optimal output prediction result of the virtual power plant, the method further includes: Based on the optimal output prediction results, dispatch control instructions for the virtual power plant are generated. Based on the scheduling and control instructions, the power allocation ratio of distributed energy resources is determined; Based on the power allocation ratio of the distributed energy source, the charging and discharging state of the energy storage system is adjusted.

10. A virtual power plant output prediction system, characterized in that, The system includes: The data processing module is used to acquire a standardized feature dataset, which is achieved by preprocessing the acquired multi-source time-series operation data of the virtual power plant. The multi-source time-series operation data includes historical output data of the virtual power plant, numerical weather forecast data, power grid dispatch instruction data, and market electricity price data. The model building module is used to construct a lightweight time-series prediction model based on a simplified gated loop unit and a time-series attention module, wherein the time-series attention module is embedded in the back end of the simplified gated loop unit; the standardized feature dataset is input into the lightweight time-series prediction model to obtain the initial output prediction results of the virtual power plant; The error correction module is used to identify the current operating condition of the virtual power plant based on historical prediction error patterns. Based on the operating condition, a differentiated correction strategy is used to perform hierarchical adaptive deviation correction on the initial output prediction result to obtain the optimal output prediction result of the virtual power plant. The differentiated correction strategy includes: when the operating condition is in a steady state, fixed compensation is performed using historical average deviation; when the operating condition is under sudden disturbance, dynamic fitting correction is performed by matching similar error curves in the historical error feature library.

Images