A virtual power plant response effect evaluation method and system

Abstract

A virtual power plant response effect evaluation method and system, historical data are collected and time sequence characteristic sequences and context characteristic vectors containing operation scene characteristics are constructed; the two are input into a context attention-based deep time sequence network CA-DTN model, an original baseline load curve is predicted, the original baseline is dynamically corrected based on actual load data of a response event day, an approved baseline is generated, the approved baseline is analyzed and judged to determine whether it belongs to an abnormal fluctuation day, when it is determined to be abnormal, the approved baseline is subjected to secondary dynamic checking, including model performance drift detection and baseline backtracking consistency analysis, and the CA-DTN model is optimized according to the checking result, and the final approved baseline is output; the actual load curve is compared with the final approved baseline, core indexes such as response amount and response completion rate are calculated, and a visual evaluation report is generated. The corresponding system includes function modules for implementing the above steps. The CA-DTN model significantly improves the prediction accuracy and market adaptability.

Description

Technical Field

[0001] This invention belongs to the technical fields of power system operation, smart grid and power market, and particularly relates to a method and system for evaluating the response effect of a virtual power plant. Background Technology

[0002] In recent years, the penetration rate of distributed renewable energy, represented by wind power and photovoltaics, has increased rapidly. However, its inherent intermittency and volatility pose significant challenges to the safe and stable operation of the power system. As a key technology that aggregates distributed energy resources such as distributed power sources, energy storage systems, controllable loads, and electric vehicles, the core value of virtual power plants lies in their "adjustable and controllable" capabilities, which enable them to participate in various electricity market services (such as frequency regulation and peak shaving in the ancillary services market, electricity trading in the spot market, and demand response).

[0003] Currently, the evaluation of virtual power plant response performance mainly relies on baseline load to calculate the response, i.e., the difference between the actual load and the baseline load. However, existing technologies have the following key drawbacks: 1. Ignoring differences in market environment: For example, patent document CN120598275A inputs processed multi-source data from virtual power plants into a pre-established LSTM network model for training, outputting a virtual power plant power baseline prediction value, and then scheduling the virtual power plant based on the virtual power plant power baseline prediction value and the set actual value. Most of the baseline prediction models in the aforementioned existing technologies construct a "general" baseline, failing to fully consider the significant differences in baseline definition, calculation period, accuracy requirements, and load patterns when virtual power plants participate in different electricity markets. For example, the frequency regulation baseline of the ancillary service market requires real-time high-frequency forecasts at the minute or even second level, while the baseline of the demand response market focuses more on load patterns during specific event periods (such as 2-4 hours). Existing methods lack explicit modeling and adaptive capabilities for this market context information.

[0004] 2. Poor model interpretability and insufficient feature utilization: For example, the patent document with publication number CN116933109A uses cluster analysis of historical load data to classify virtual power plants into different types according to their operating characteristics and establishes a baseline load forecasting model. Although deep learning models (such as LSTM) have a strong ability to capture time-series dependencies, their "black box" nature makes the model's decision-making process difficult to interpret, and it is impossible to clearly know which features (such as specific weather, historical load during the same period, and market conditions) play a key role at a specific forecast time. Global contextual information (such as market type, major event labels, and seasonal cycles) is often simply concatenated into the input without being effectively structured and assigned different levels of importance.

[0005] 3. Lack of dynamic correction and anti-interference mechanisms: For example, the LSTM model with publication number CN120598275A is usually a static model, and its parameters are fixed once training is completed. When faced with complex real-world situations such as temporary user maintenance, sudden changes in holiday patterns, and abnormal load fluctuations (such as production plan adjustments for non-market reasons), the lack of effective online correction, data cleaning, and anti-interference mechanisms leads to serious distortion of baseline prediction results and evaluation failure.

[0006] 4. Lack of the concept of "optimal controllable baseline": The baseline is not only an extrapolation of historical data, but should also reflect the "optimal operating state" that virtual power plant operators can achieve under normal circumstances, which is relatively stable and cost-effective. Most existing research focuses on minimizing prediction errors and rarely actively defines and guides the model to learn this "optimal controllable" baseline standard from the perspective of operation optimization.

[0007] 5. Inadequate verification and assurance system: Existing evaluation processes are mostly open-loop systems, lacking pre-emptive control over data quality, post-event verification of evaluation results, and long-term monitoring and evolution mechanisms for model performance. When disputes arise, the lack of transparent and reliable traceability and review methods leads to high user complaint rates and frequent market disputes.

[0008] Therefore, there is an urgent need for an innovative and systematic virtual power plant response performance evaluation scheme that can achieve high-precision prediction of baseline load, intelligent dynamic correction, and full-process reliable verification in complex and ever-changing market and operating environments, thereby providing a solid technical guarantee for the fair and efficient operation of the electricity market. Summary of the Invention

[0009] The primary objective of this invention is to overcome the aforementioned deficiencies of the prior art and provide a method and system for evaluating the response performance of a virtual power plant. This invention aims to achieve the following specific objectives: A high-precision baseline load forecasting model that can adapt to different electricity market environments (ancillary services, spot market, demand response) is constructed to significantly improve forecasting accuracy, especially during critical peak load periods. Design a baseline dynamic intelligent correction mechanism based on user electricity consumption behavior analysis, which can effectively identify and isolate the impact of non-market factors on the load, and ensure the purity and fairness of response quantity calculation.

[0010] Establish a multi-layered protection and closed-loop verification system covering the entire chain of "prevention, correction, and verification" to ensure data quality from the source, strengthen anti-interference capabilities throughout the process, and continuously monitor and optimize the stability and reliability of long-term evaluation results.

[0011] The introduction of the guiding concept of "optimal controllable baseline" enables baseline prediction to not only fit history, but also approach the ideal benchmark of virtual power plants under optimized operation, thereby enhancing the practicality and guiding significance of the baseline.

[0012] To develop a complete, feasible, and efficient technical implementation plan, providing transparent, reliable, and automated assessment tools and decision support for power trading institutions, virtual power plant operators, and relevant regulatory authorities.

[0013] To achieve the above objectives, this invention proposes a virtual power plant response performance evaluation method and system that is "accurate prediction as the core, dynamic correction as the wings, and multiple verifications as the shield." The core idea of ​​this method and system lies in constructing a multi-level, highly adaptive, closed-loop feedback intelligent evaluation framework.

[0014] The present invention adopts the following technical solution.

[0015] This invention proposes a method for evaluating the response effect of a virtual power plant, comprising: Step 1: Collect historical aggregated power data, meteorological data, calendar information, and market transaction records of the target virtual power plant, construct a time-series feature sequence and a context feature vector, and perform preprocessing; the time-series feature sequence includes a fluctuating continuous time series, and the context feature vector includes static information of the operating scenario features; Step 2: Input the preprocessed temporal feature sequence and contextual feature vector into the constructed deep temporal network CA-DTN model to predict the original baseline load curve for the response event day; Step 3: Obtain the actual load data and event information for the response event day, dynamically correct the original baseline load curve, and generate the approved baseline; Step 4: Based on the preset multi-dimensional load fluctuation rationality judgment rule set, the approved baseline is judged to determine whether it belongs to an abnormal fluctuation day; Step 5: When an abnormal fluctuation day is identified, a second dynamic verification is performed on the approved baseline, including model performance drift detection and baseline backtracking consistency analysis; based on the results of the second dynamic verification, the CA-DTN model is retrained, and the final approved baseline load curve is output using the optimized CA-DTN model. Step 6: Compare the actual load curve of the response event day with the final approved baseline load curve, calculate the actual response amount, response completion rate and regulation accuracy index of the virtual power plant, and generate a visual evaluation report that includes curve comparison, index statistics and anomaly description.

[0016] More preferably, in step 2, the CA-DTN model includes an input layer, a context branch, a temporal branch, and an output layer; The input layer receives temporal feature sequences and context feature vectors in parallel. The context branch includes an embedding layer and a context attention module, which ultimately outputs a weighted context vector. The temporal branch includes a vector augmentation layer, a deep temporal encoder, and a temporal attention module, which ultimately outputs a weighted temporal vector. The output layer concatenates the weighted context vector and the weighted time series vector and then obtains the baseline load curve through a fully connected structure.

[0017] More preferably, in step 2, the context branch includes an embedding layer and a context attention module: In the embedding layer, categorical features in the context feature vector are embedded and encoded, while numerical features are standardized to form a dense vector.

[0018] In the context attention module, The features in the dataset are divided into at least two feature groups based on their categories. Contextual attention weights are calculated for each feature group using independent multilayer perceptron and softmax layers. Based on the contextual attention weights of each feature group, the... The corresponding feature vectors are weighted, and then the weighted features in all feature groups are summed to generate a weighted context vector.

[0019] More preferably, in step 2, the temporal branch includes a vector enhancement layer, a deep temporal encoder, and a temporal attention module, specifically implemented as follows: In the vector enhancement layer, the CA-DTN model also includes a deep temporal network module, which receives temporal feature sequences. With the context representation vector The specific processing procedure is as follows: Temporal feature embedding: embedding temporal feature sequences Layer mapping to a high-dimensional space yields temporal feature embeddings. The weighted context vector output from the context branch is broadcast to each time step and concatenated with the temporal feature embedding to form an enhanced temporal input; The basic architecture of a deep temporal encoder is either BiLSTM or GRU, which encodes the enhanced temporal input: a multi-layer bidirectional LSTM is used to process the enhanced temporal input, obtain the forward and backward hidden states, and concatenate the states at the corresponding time steps to obtain the temporal representation. ; In the temporal attention module, a similarity score is calculated between the trainable query vector and each hidden state in H, and the temporal attention weights are obtained by normalizing the similarity scores. Then, H is weighted and summed based on the temporal attention weights to generate a weighted temporal vector.

[0020] More preferably, in step 3, the dynamic correction method is a three-stage correction method, specifically: The first phase involves selecting non-execution periods on the response event day that are not affected by market orders. Calculate the average relative deviation between the actual load and the original baseline load curve during this period. ;like Less than the preset threshold If the result is satisfactory, proceed to the third stage of residual fitting to generate the final baseline; otherwise, after the second stage of preliminary offset correction, input the corrected baseline load curve into the third stage.

[0021] The second stage involves fitting and correcting the original baseline load curve: During the execution period of the response event Using average relative offset The original baseline load curve is weighted to obtain the corrected baseline load curve; The third stage involves calculating non-execution periods. Within this, the actual load is compared with the original baseline load curve or the corrected baseline load curve obtained after the second stage. sequence ; Spline interpolation or Gaussian process regression methods are used to analyze discrete residual points. By fitting the data, a smoothing correction function defined over the entire time period is obtained.

[0022] Will The approved baseline is generated by adding the original baseline load curve or the modified baseline load curve obtained in the second stage.

[0023] More preferably, in step 4, the multi-dimensional load fluctuation rationality assessment rule set includes judgment rules for five dimensions: total load fluctuation, abnormal curve shape, peak / trough detection, and abnormal rate of change. Total fluctuation dimension: Calculate the total electricity consumption on the target day. If its daily average value for the same period in history is The absolute difference exceeds 100 times the historical standard deviation If so, an exception flag is triggered, where These are preset coefficients; Anomaly dimension of curve shape: Calculate the target daily load curve L and the historical typical daily load curve. Dynamic time warping distance If this distance exceeds the historical normal daily DTW distance distribution If the quantile is reached, an exception flag is triggered; Peak / trough detection dimension: Employs a sliding window statistical method to detect peaks / troughs that consistently exceed the mean within the window. If the number of local extreme points exceeds a threshold, or if the isolated forest algorithm determines that the daily curve is an overall outlier, an anomaly marker is triggered. Anomaly dimension of rate of change: Calculate the point-by-point difference ΔL(t) of the load, and count whether its absolute value exceeds the historical normal difference range. The proportion, if the proportion exceeds a preset threshold If so, an exception flag will be triggered.

[0024] More preferably, in step 4, the specific method for determining whether a day belongs to an abnormal fluctuation is as follows: Based on the rule set for judging the rationality of load fluctuations in a multi-dimensional manner, the judgment result of each rule is generated for the approved baseline. The judgment results are then weighted and voted on to obtain a weighted total score for anomalies. If the weighted abnormal total score exceeds the set total score threshold, it is determined to be an abnormal fluctuation day.

[0025] More preferably, in step 5, the secondary dynamic verification includes model performance drift detection and baseline backtracking consistency analysis, and the implementation method includes: Model performance drift detection uses the latest historical load data for a set period as a new test set to recalculate the prediction accuracy indicators of the current CA-DTN model, including the mean absolute percentage error (MAPE). Statistical process control methods are used, with the center line set as the mean of the mean absolute percentage error over a set historical normal period, and the upper control limit set as the mean plus three standard deviations. If the mean absolute percentage error for a consecutive set number of days is higher than the upper control limit, or if the mean absolute percentage error shows a statistically significant monotonically increasing trend, then the model is determined to have experienced performance drift. Baseline retrospective consistency analysis selects a set of dates for which complete historical actual load data has been obtained and previously assessed. Using the latest CA-DTN model, the original baseline is re-predicted for each of these dates, and the dynamic correction in step 3 is performed to obtain the newly calculated approved baseline. ;Will The approved baseline used in the original assessment on that date A point-by-point comparison is performed, and quantitative indicators are calculated, including the longest duration during which the point-by-point relative error continuously exceeds the set error threshold, the daily average relative error for the whole day, and the average absolute error within the response period. If the longest duration is not less than the set time threshold, or the daily average relative error for the whole day is greater than the set threshold, or the average absolute error within the response period is greater than the set threshold of the baseline rated capacity, then a systematic difference is determined to have occurred, triggering the model retraining of the CA-DTN model.

[0026] More preferably, in step 5, the CA-DTN model is optimized based on the results of the secondary dynamic verification, and the final approved baseline load curve is output using the optimized CA-DTN model. Specific steps include: Extract all normal day samples confirmed in step 4 from historical data, take the data of the most recent set number of days in chronological order, and construct the time series feature sequence and context feature vector according to the method in step 1. The above samples are divided into training set, validation set and test set; Keep the CA-DTN model architecture exactly the same as in step 2, and retrain it on the training set; Evaluate the MAPE and RMSE metrics of the retrained CA-DTN model on the test set. If the MAPE on the test set decreases by more than a set threshold compared to the model before drift detection is triggered, the model is automatically replaced; otherwise, the original model is retained and an alarm is recorded.

[0027] This invention also proposes a virtual power plant response performance evaluation system, including a data and feature extraction module, a baseline prediction module, a dynamic correction module, a fluctuation day judgment module, a secondary dynamic verification and retraining module, and a performance evaluation module. The data and feature extraction module collects historical aggregated power data, meteorological data, calendar information, and market transaction records of the target virtual power plant, constructs a time-series feature sequence and a context feature vector, and performs preprocessing; the time-series feature sequence includes a fluctuating continuous time series, and the context feature vector includes static information of the operating scenario features; The baseline prediction module inputs the preprocessed temporal feature sequence and context feature vector into the constructed deep temporal network CA-DTN model to predict the original baseline load curve for the response event day. The dynamic correction module acquires the actual load data and event information for the response event day, dynamically corrects the original baseline load curve, and generates an approved baseline. The fluctuation day judgment module, based on a preset multi-dimensional load fluctuation rationality judgment rule set, judges the approved baseline to determine whether it belongs to an abnormal fluctuation day; The secondary dynamic verification and retraining module performs a secondary dynamic verification of the approved baseline when an abnormal fluctuation day is identified. This includes model performance drift detection and baseline backtracking consistency analysis. Based on the results of the secondary dynamic verification, the CA-DTN model is retrained, and the optimized CA-DTN model is used to output the final approved baseline load curve. The effect evaluation module compares the actual load curve of the response event day with the final approved baseline load curve, calculates the actual response volume, response completion rate and regulation accuracy of the virtual power plant, and generates a visual evaluation report that includes curve comparison, indicator statistics and anomaly descriptions.

[0028] The present invention also proposes a terminal, including a processor and a storage medium: The storage medium is used to store instructions; The processor is used to perform the steps of the above method according to the instructions.

[0029] The present invention also proposes a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the above-described method.

[0030] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. Significantly improved forecast accuracy and enhanced market adaptability: The CA-DTN model explicitly models market differences through a contextual attention mechanism, making baseline forecasts strongly correlated with specific market scenarios. Experiments show that in spot market scenarios, its MAPE is reduced by approximately 20% compared to the standard LSTM model, and peak load forecasting errors are significantly reduced. The model can adaptively switch from minute-level frequency modulation baselines to day-ahead 96-point baseline forecasts, demonstrating strong versatility.

[0031] 2. Significantly Improved Fairness in Assessment: The dynamic correction mechanism effectively isolates load bias caused by non-market factors, making the calculation of response quantities more "pure." Simulation tests show that after adopting the complete method chain (prediction + correction + verification), the deviation in response quantity assessment can be reduced from ±8.2% of the traditional method to within ±3.5%, greatly reducing market disputes caused by baseline injustice.

[0032] 3. Enhanced System Robustness and Reliability: A triple-checking and safeguarding mechanism forms a robust "safety net." Maintenance requests purify data at the source, fluctuation analysis automatically filters out abnormal days, and secondary checks continuously monitor system health. This ensures the evaluation system maintains stable and reliable output even in the face of real-world interference such as data anomalies and sudden changes in user behavior. In actual pilot testing, the user complaint rate regarding evaluation results decreased by over 70%.

[0033] 4. Transparent and credible assessment process with improved efficiency: The entire assessment process, including algorithm logic, parameter corrections, and verification records, is digitally documented and visualized, enhancing transparency and interpretability. Simultaneously, the highly automated process reduces the assessment cycle, which previously required significant manual intervention, from days to hours or even minutes, greatly improving the efficiency of electricity market operation and settlement.

[0034] 5. Possesses continuous evolution capability: The secondary dynamic verification mechanism and expert feedback closed loop make the system no longer a static tool, but an intelligent system that can continuously optimize and adapt itself as market rules change, user behavior evolves, and data accumulates, ensuring the long-term vitality of the technical solution. Attached Figure Description

[0035] Figure 1 This is a flowchart of a virtual power plant response effect evaluation method according to the present invention; Figure 2 This is a flowchart of the virtual power plant response effect evaluation method in this invention; Figure 3 This is a detailed architectural diagram of the CA-DTN model in this invention; Figure 4 This is a flowchart of the baseline load dynamic correction algorithm based on user electricity consumption characteristics in this invention; Figure 5 This is a schematic diagram of the secondary dynamic verification process in this invention. Detailed Implementation

[0036] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of this invention. The embodiments described in this application are merely some embodiments of this invention, and not all embodiments. Based on the spirit of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the protection scope of this invention.

[0037] like Figure 1 As shown, this invention proposes a method for evaluating the response effect of a virtual power plant, which specifically includes the following steps: Step 1: Collect historical aggregated power data, meteorological data, calendar information, and market transaction records of the target virtual power plant, construct a time-series feature sequence and a context feature vector, and perform preprocessing; the time-series feature sequence includes a fluctuating continuous time series, and the context feature vector includes static information of the operating scenario features; The CA-DTN model includes an input layer, a context branch, a temporal branch, and an output layer. The input layer receives temporal feature sequences and context feature vectors in parallel. The context branch includes an embedding layer and a context attention module, which ultimately outputs a weighted context vector. The temporal branch includes a vector augmentation layer, a deep temporal encoder, and a temporal attention module, which ultimately outputs a weighted temporal vector. The output layer concatenates the weighted context vector and the weighted time series vector and then obtains the baseline load curve through a fully connected structure.

[0038] The context branch includes the embedding layer and the context attention module: In the embedding layer, categorical features in the context feature vector are embedded and encoded, while numerical features are standardized to form a dense vector.

[0039] In the context attention module, The features in the dataset are divided into at least two feature groups based on their categories. Contextual attention weights are calculated for each feature group using independent multilayer perceptron and softmax layers. Based on the contextual attention weights of each feature group, the... The corresponding feature subvectors are weighted, and then the weighted features in all feature groups are summed to generate a weighted context vector.

[0040] The temporal branch includes a vector augmentation layer, a deep temporal encoder, and a temporal attention module. The specific implementation method is as follows: In the vector enhancement layer, the CA-DTN model also includes a deep temporal network module, which receives temporal feature sequences. With the context representation vector The specific processing procedure is as follows: Temporal feature embedding: embedding temporal feature sequences Layer mapping to a high-dimensional space yields temporal feature embeddings. The weighted context vector output from the context branch is broadcast to each time step and concatenated with the temporal feature embedding to form an enhanced temporal input; The basic architecture of a deep temporal encoder is either BiLSTM or GRU, which encodes the enhanced temporal input: a multi-layer bidirectional LSTM is used to process the enhanced temporal input, obtain the forward and backward hidden states, and concatenate the states at the corresponding time steps to obtain the temporal representation. ; In the temporal attention module, a similarity score is calculated between the trainable query vector and each hidden state in H, and the temporal attention weights are obtained by normalizing the similarity scores. The weighted sum of H is then calculated based on the temporal attention weights to generate a weighted temporal vector.

[0041] Step 2: Input the preprocessed temporal feature sequence and contextual feature vector into the constructed deep temporal network CA-DTN model to predict the original baseline load curve for the response event day; Step 3: Obtain the actual load data and event information for the response event day, dynamically correct the original baseline load curve, and generate the approved baseline; The dynamic correction method is a three-stage correction method, specifically: The first phase involves selecting non-execution periods on the response event day that are not affected by market orders. Calculate the actual load during this period. Average relative deviation of baseline load curve ;like Less than the preset threshold If the result is satisfactory, proceed to the third stage of residual fitting to generate the final baseline; otherwise, after the second stage of preliminary offset correction, input the corrected baseline load curve into the third stage.

[0042] The second stage involves fitting and correcting the original baseline load curve: During the execution period of the response event Using average relative offset The original baseline load curve is weighted to obtain the corrected baseline load curve; The third stage involves calculating non-execution periods. Within this, the actual load is compared with the original baseline load curve or the corrected baseline load curve obtained after the second stage. sequence ; Spline interpolation or Gaussian process regression methods are used to analyze discrete residual points. By fitting the data, a smoothing correction function defined over the entire time period is obtained.

[0043] Will The approved baseline is generated by adding the original baseline load curve or the modified baseline load curve obtained in the second stage.

[0044] Step 4: Based on the preset multi-dimensional load fluctuation rationality judgment rule set, the approved baseline is judged to determine whether it belongs to an abnormal fluctuation day; The multi-dimensional load fluctuation rationality assessment rule set includes judgment rules based on five dimensions: total load fluctuation, abnormal curve shape, peak / trough detection, and abnormal rate of change. Total fluctuation dimension: Calculate the total electricity consumption on the target day. If its daily average value for the same period in history is The absolute difference exceeds 100 times the historical standard deviation If so, an exception flag is triggered, where These are preset coefficients; Anomaly dimension of curve shape: Calculate the target daily load curve L and the historical typical daily load curve. Dynamic time warping distance If this distance exceeds the historical normal daily DTW distance distribution If the quantile is reached, an exception flag is triggered; Peak / trough detection dimension: Employs a sliding window statistical method to detect peaks / troughs that consistently exceed the mean within the window. If the number of local extreme points exceeds a threshold, or if the isolated forest algorithm determines that the daily curve is an overall outlier, an anomaly marker is triggered. Anomaly dimension of rate of change: Calculate the point-by-point difference ΔL(t) of the load, and count whether its absolute value exceeds the historical normal difference range. The proportion, if the proportion exceeds a preset threshold If so, an exception flag will be triggered.

[0045] The specific method for determining whether a day belongs to abnormal fluctuations is as follows: Based on the rule set for judging the rationality of load fluctuations in a multi-dimensional manner, the judgment result of each rule is generated for the approved baseline. The judgment results are then weighted and voted on to obtain a weighted total score for anomalies. If the weighted abnormal total score exceeds the set total score threshold, it is determined to be an abnormal fluctuation day.

[0046] Step 5: When an abnormal fluctuation day is identified, a second dynamic verification is performed on the approved baseline, including model performance drift detection and baseline backtracking consistency analysis; based on the results of the second dynamic verification, the CA-DTN model is retrained, and the final approved baseline load curve is output using the optimized CA-DTN model. Secondary dynamic verification, including model performance drift detection and baseline backtracking consistency analysis, is implemented through the following methods: Model performance drift detection uses the latest historical load data for a set period as a new test set to recalculate the prediction accuracy indicators of the current CA-DTN model, including the mean absolute percentage error (MAPE). Statistical process control methods are used, with the center line set as the mean of the mean absolute percentage error over a set historical normal period, and the upper control limit set as the mean plus three standard deviations. If the mean absolute percentage error for a consecutive set number of days is higher than the upper control limit, or if the mean absolute percentage error shows a statistically significant monotonically increasing trend, then the model is determined to have experienced performance drift. Baseline retrospective consistency analysis selects a set of dates for which complete historical actual load data has been obtained and previously assessed. Using the latest CA-DTN model, the original baseline is re-predicted for each of these dates, and the dynamic correction in step 3 is performed to obtain the newly calculated approved baseline. ;Will The approved baseline used in the original assessment on that date A point-by-point comparison is performed, and quantitative indicators are calculated, including the longest duration during which the point-by-point relative error continuously exceeds the set error threshold, the daily average relative error for the whole day, and the average absolute error within the response period. If the longest duration is not less than the set time threshold, or the daily average relative error for the whole day is greater than the set threshold, or the average absolute error within the response period is greater than the set threshold of the baseline rated capacity, then a systematic difference is determined to have occurred, triggering the model retraining of the CA-DTN model.

[0047] The final approved baseline load curve is output using the optimized CA-DTN model. The specific steps include: Extract all normal day samples confirmed in step 4 from historical data, take the data of the most recent set number of days in chronological order, and construct the time series feature sequence and context feature vector according to the method in step 1. The above samples are divided into training set, validation set and test set; Keep the CA-DTN model architecture exactly the same as in step 2, and retrain it on the training set; Evaluate the MAPE and RMSE metrics of the retrained CA-DTN model on the test set. If the MAPE on the test set decreases by more than a set threshold compared to the model before drift detection is triggered, the model is automatically replaced; otherwise, the original model is retained and an alarm is recorded.

[0048] Step 6: Compare the actual load curve of the response event day with the final approved baseline load curve, calculate the actual response amount, response completion rate and regulation accuracy index of the virtual power plant, and generate a visual evaluation report that includes curve comparison, index statistics and anomaly description.

[0049] This invention also proposes a virtual power plant response performance evaluation system, including a data and feature extraction module, a baseline prediction module, a dynamic correction module, a fluctuation day judgment module, a secondary dynamic verification and retraining module, and a performance evaluation module. The data and feature extraction module collects historical aggregated power data, meteorological data, calendar information, and market transaction records of the target virtual power plant, constructs a time-series feature sequence and a context feature vector, and performs preprocessing; the time-series feature sequence includes a fluctuating continuous time series, and the context feature vector includes static information of the operating scenario features; The baseline prediction module inputs the preprocessed temporal feature sequence and context feature vector into the constructed deep temporal network CA-DTN model to predict the original baseline load curve for the response event day. The dynamic correction module acquires the actual load data and event information for the response event day, dynamically corrects the original baseline load curve, and generates an approved baseline. The fluctuation day judgment module, based on a preset multi-dimensional load fluctuation rationality judgment rule set, judges the approved baseline to determine whether it belongs to an abnormal fluctuation day; The secondary dynamic verification and retraining module performs a secondary dynamic verification of the approved baseline when an abnormal fluctuation day is identified. This includes model performance drift detection and baseline backtracking consistency analysis. Based on the results of the secondary dynamic verification, the CA-DTN model is retrained, and the optimized CA-DTN model is used to output the final approved baseline load curve. The effect evaluation module compares the actual load curve of the response event day with the final approved baseline load curve, calculates the actual response volume, response completion rate and regulation accuracy of the virtual power plant, and generates a visual evaluation report that includes curve comparison, indicator statistics and anomaly descriptions.

[0050] The present invention also proposes a terminal, including a processor and a storage medium: The storage medium is used to store instructions; The processor is used to perform the steps of the above method according to the instructions.

[0051] The present invention also proposes a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the above-described method.

[0052] Example 1 To address the shortcomings of existing technologies, this invention proposes a method and system for evaluating the response performance of virtual power plants, providing a fair and accurate evaluation basis for virtual power plants participating in diversified electricity market transactions. See also... Figure 2 As shown, the virtual power plant response effect evaluation method disclosed in this invention includes the following steps: Step 1: Multi-source data fusion and preprocessing. Integrate historical and real-time data from metering and data acquisition systems, electricity market trading platforms, meteorological service systems, and user reporting terminals. After standardizing the collected data through cleaning, alignment, and normalization, construct time-series feature sequences (such as continuous time series of load, temperature, and other fluctuations over the past N days) and contextual feature vectors (such as static information about the market type, date attributes, season, and other operational scenario characteristics of the target prediction day).

[0053] The standardized processing procedures include: Missing value handling: For random missing values, time series linear interpolation is used; for continuous long-term missing values ​​(>2 hours), they are marked and filled using daily data of the same type in the next period.

[0054] Outlier handling: A combination of rolling statistics and isolation forest is used. First, a sliding window (3σ principle) is used to remove instantaneous spikes, and then an unsupervised isolation forest algorithm is used to detect and remove abnormal load curves throughout the day.

[0055] Normalization: "Max-Min normalization" is performed on all numerical features, mapping the values ​​to the [0,1] interval to accelerate model convergence. Formula: .

[0056] Sequence construction: For each prediction task, construct samples as follows .in, C is the feature sequence of the backtracking window (e.g., 96 points per day for the past 10 days), C is the context vector of the prediction day, and Y is the prediction target (load of 96 points in the next 24 hours).

[0057] Step 2: CA-DTN Prediction Model Construction. The preprocessed temporal feature sequence and contextual feature vector are input into the core prediction engine—a deep temporal network based on contextual attention (CA-DTN). This model, through an innovative grouped attention mechanism, deeply mines the multi-scale temporal dependencies of historical loads and dynamically adjusts the "attention focus" according to the context of the current prediction task (especially the market environment), outputting a high-precision original baseline load curve that conforms to specific market rules. See also... Figure 3 As shown, the specific structure of CA-DTN includes an input layer, a context branch, a timing branch, and an output layer: In the input layer, the model receives time-series feature sequences and contextual feature vectors in parallel to simultaneously capture the dynamic changes of the time series and the global static information of the predicted scene. Input1: Temporal feature sequence

[0058] in, For historical time steps (e.g., 96 points per day for the past 10 days, totaling 960 points; or a downsampled representation of a longer sequence); The feature dimensions for each time point include: (e.g., the total historical power of the virtual power plant (kW), meteorological data including temperature, humidity and wind speed, cycle time features such as the number of hours in a day and the day of the week, lag features such as the load values ​​of the previous day or the previous 7 days, etc.).

[0059] Input2: Context feature vector

[0060] in, The context feature vector is a digitized vector constructed by one-hot encoding or embedding encoding of static information. Specifically, various categorical static information describing the prediction scenario (such as market type, date type) is converted into a string of 0s and 1s with clear physical meaning through one-hot encoding or other methods. Simultaneously, some numerical or binary information (such as month, whether there are special events) is standardized or directly encoded. All these encoded numbers are then concatenated to form a complete context feature vector, which is then input into the model.

[0061] A context feature vector is a configurable and scalable set of features that effectively encodes and inputs static or quasi-static information about the prediction scenario itself—influenced by the prediction target but which cannot be directly represented by time series data—into the model. Features can be flexibly defined, added, or removed based on the specific virtual power plant aggregation object, the type of market involved (e.g., frequency regulation, peak shaving, demand response), and the available data.

[0062] The context branch includes the embedding layer and the context attention module: Embedding layer, for Categorical features are embedded, while numerical features are directly preserved or standardized to form dense vectors. .

[0063] In the context attention module, dense vectors The raw attention score is obtained after inputting into a multilayer perceptron (MLP). ,Will Inputting the softmax layer yields the context attention weights. :

[0064]

[0065] in, Dimensions and The sub-feature dimension or grouping dimension is related, and grouping attention weights can be calculated. For example, the context attention weights can be calculated by inputting the corresponding multilayer perceptron and Softmax layer for "market feature group" and "date feature group" respectively.

[0066] The context attention weights are weighted and summed with the dense vector to obtain the weighted context vector. :

[0067] This vector contains the most important contextual information for the current prediction task. In the table below, i represents the position of an element.

[0068] The specific structure of the temporal branch includes a vector augmentation layer, a deep temporal encoder, and a temporal attention module: Vector enhancement layer, which transforms temporal feature sequences By mapping to a high-dimensional space through a linear layer, the temporal feature embedding feature is obtained. .

[0069] The weighted context vector output by the context branch Broadcast (copy) to each time step, with By splicing, an enhanced timing input is formed. This allows BiLSTM to be aware of the global context at every time step.

[0070] The basic architecture of a deep temporal encoder is either BiLSTM or GRU, which encodes the enhanced temporal input: a multi-layer bidirectional LSTM is used to process the enhanced temporal input, obtain the forward and backward hidden states, and concatenate the states at the corresponding time steps to obtain a rich temporal representation. .

[0071] The temporal attention module focuses on the most critical historical moments for predicting future periods: Using a trainable query vector (Or use the last hidden state) and temporal representation Each of them Calculate similarity score .in, For trainable weight vectors, , is a trainable weight matrix; Normalized similarity scores yield temporal attention weights:

[0072] Generate weighted temporal vectors based on temporal attention weights: This vector concentrates the most relevant information from the historical sequence.

[0073] In the output layer, the weighted time series vectors are... and weighted context vector Concatenate the vectors to form the final joint representation vector. .

[0074] Joint representation vector The baseline load curve is obtained through a multilayer perceptron (MLP), i.e., a fully connected structure. Baseline load curve include Points in time:

[0075] in, These are the weight matrix and bias of the hidden layer in a multilayer perceptron, respectively. These are the weight matrix and bias of the output layer in a multilayer perceptron, respectively.

[0076] The architecture of the CA-DTN model enables market adaptation: market type adaptation is mainly achieved through contextual attention and differential training.

[0077] Among the contextual features, market type is one of the key features. The model learns to focus on different historical patterns and external factors under different market types.

[0078] During training data preparation, data from different market scenarios are labeled accordingly. The model automatically learns and invokes the "knowledge" most relevant to the current market type through a contextual attention mechanism.

[0079] Different output heads can be set for different markets (Shared-Bottom multi-task learning framework), but this solution uses a unified output head and relies on contextual features for conditional generation to enhance model flexibility and shared representation learning.

[0080] Training a CA-DTN model requires a carefully designed total loss function to simultaneously satisfy the requirements of accuracy, controllability, and stationarity.

[0081] The total loss function is composed of three weighted components: mean squared error loss, smoothness loss, and controllability guidance loss.

[0082] in, The weighting coefficient is determined based on experimental performance.

[0083] Mean Squared Error Loss (MSE): Ensures the accuracy of the basic prediction.

[0084]

[0085] in, For the true value, These are predicted values.

[0086] Smoothness Loss: Encourages a smoother predicted baseline curve to avoid unnecessary sharp fluctuations, which is more in line with the characteristics of "normal" operation and is beneficial to power grid dispatch.

[0087]

[0088] Controllability Loss: This is crucial for achieving the "optimal controllable baseline." A reference target is required. This represents the "ideal" operating curve of a virtual power plant under conditions of no market interference. This curve can be obtained through simulation using offline optimization models (such as those aiming for the lowest operating cost and the most stable load) or defined by expert rules.

[0089]

[0090] By minimizing this, the model's predictions are guided toward the "optimal controllable" state. The calculation can be based on historical data and typical virtual power plant operating constraints.

[0091] Step 3: Dynamic Baseline Load Correction. After obtaining the actual load data for the response event day, an intelligent correction algorithm is executed. Its core logic is based on the assumption that the impact of non-market factors (such as sudden weather changes or spontaneous production adjustments) on electricity consumption behavior is consistent or inferable between the response and non-response periods. By analyzing the deviation between the actual load and the predicted baseline during the non-response period, a smoothing correction function is constructed to calibrate and fit the original baseline load curve, generating an approved baseline that more closely resembles the actual "assumed no response" scenario.

[0092] Specifically, it is assumed that changes in electricity consumption behavior caused by non-market factors are consistent or inferable during and outside the response event execution period. Required input data includes: CA-DTN Predicted Baseline Load Curve (All day).

[0093] Actual load on the response event day .

[0094] Response event information: Execution period And market type.

[0095] Based on the above input data, corrections are made using a dynamic correction algorithm. See also... Figure 4 As shown, the correction process consists of three stages: The first stage involves determining the overall offset caused by non-market factors and calculating the offset during non-execution periods: Data Extraction: On the event day, collect data from periods unaffected by market response, typically several hours before and after the execution period (e.g., 2 hours before the event starts and 2 hours after it ends). This is recorded as the non-execution period set. .

[0096] Offset calculation: Calculate the actual load during this period. Compared with the predicted original baseline load curve Average relative offset between .

[0097]

[0098] in, Points counted during non-execution periods. Set threshold. (like <5%), if Then, the second stage of preliminary offset correction will be performed; if If no significant systematic shift is found, the process can proceed directly to the third stage of residual fitting to generate the approved baseline.

[0099] In the second stage, to generate a smooth and reasonable final baseline curve, the original baseline load curve is modified. Perform fit correction: Initial offset correction: If A preliminary proportional correction is performed on the baseline throughout the day to obtain the corrected baseline load curve. :

[0100] The third stage involves generating the approval baseline: Non-execution period residual modeling: Calculating non-execution periods Internally, actual load and residual sequence .

[0101] Constructing the correction function: Using spline interpolation or Gaussian process regression on discrete residual points By fitting the data, a smooth correction function defined throughout the day is obtained. This function is in It accurately crosses the residual point and provides reasonable interpolation estimates for other time periods (especially the execution period).

[0102] Generate approved baseline: The approved baseline is obtained by adding it to either the original baseline load curve or the modified baseline load curve obtained in the second stage. . The design ensures a smooth transition of corrections, avoiding unreasonable jumps at the execution time boundary.

[0103] Step 4: Load Fluctuation Reasonableness Assessment. To ensure the quality of data used for baseline calculation or correction reference, the system has a built-in multi-dimensional load fluctuation reasonableness assessment rule set. This rule set automatically scans historical candidate days or approved baselines and comprehensively determines whether they belong to abnormal fluctuation days (such as unreported maintenance or special activities) through multiple dimensions such as total fluctuation, curve shape, local extreme values, and rate of change.

[0104] Step 4.1: Construct a multi-dimensional set of rules for judging the rationality of load fluctuations, including judgment rules for five dimensions: total load fluctuation, abnormal curve shape, peak / trough detection, and abnormal rate of change. The objective of the total fluctuation dimension is to detect whether daily electricity consumption deviates significantly from historical normal levels. The specific rules and algorithms are as follows: Calculate the total electricity consumption for the day The daily average value for the same period in history and standard deviation Comparison: If If the condition is met, an exception flag is triggered, with the flag set to 1; otherwise, it is 0. The threshold value is... Typically, the range is 2.5 to 3.0 (based on the assumption of normal distribution, corresponding to a 99% confidence interval).

[0105] The objective of the curve anomaly dimension is to determine whether the shape of the load curve deviates significantly from the typical pattern. The specific rules and algorithms are as follows: Calculate the daily load curve sequence L and the historical typical curve. Dynamic Time Warping (DTW) distance; DTW can flexibly align the time axis and effectively measure shape differences: if If this occurs, an anomaly flag is triggered; where the distance threshold is... Take the 95th percentile of the DTW distance distribution on historical normal days; typical curve It is generated by averaging the historical normal daily load curves.

[0106] The goal of peak / valley detection is to identify unreasonable local extreme points in a curve. The specific rules and algorithms are as follows: A. Local Extremum Statistical Method: Using a sliding window, the mean within the window is calculated. and standard deviation If the load value at a certain point continuously exceeds... If the threshold coefficient is 0, it is determined to be a local extremum point; where the threshold coefficient is 0. Typically set to 3.0; Duration Points The value is usually 2 (to avoid misjudgment due to noise). B. Isolation Forest (IF) Model: Using the daily load curve as sample points, the unsupervised IF algorithm is used to detect outliers in the whole; If the number of extreme points detected by method A exceeds the threshold or method B determines it to be an outlier, an anomaly marker is triggered.

[0107] The purpose of the rate of change anomaly dimension is to detect situations where the rate of change in load power is abnormally rapid. The specific rules and algorithms are as follows: Calculate the minute-level differential of the load. The absolute value of the data exceeds the historical normal range. Proportion: If If so, an anomaly flag is triggered; where the threshold coefficient is... Typically, a value of 3.0 is used; the proportional threshold. It is usually set at 1%-5%; the historical normal range is obtained by statistical analysis of the difference data of historical normal days.

[0108] By employing a comprehensive analysis of multiple indicators, a hybrid model of "rule engine + lightweight model" is formed.

[0109]

[0110] Step 4.2, Integrated Decision-Making and Automated Processing Flow Weighted voting: Based on a multi-dimensional set of rules for assessing the rationality of load fluctuations, a judgment result is generated for each rule on the approved baseline, and a weight is assigned to each judgment result. (For example, the total fluctuation has the highest weight, followed by the abnormal shape), calculate the weighted total score of abnormality.

[0111] Automatic adjudication: Set total score threshold If the total score exceeds the threshold, the system will automatically classify the day as an "unreasonable fluctuation day" and remove it from the pool of candidate typical days calculated from the baseline.

[0112] As a preferred embodiment of the present invention, it may further include: Manual review queue: For dates with total scores near the threshold (fuzzy zone), or dates that trigger specific important rules (such as detecting extreme spikes lasting for a long period), the system places them in the "awaiting manual review" queue and highlights the abnormal characteristics for final decision-making by operations experts. The reason for setting up the manual review queue is to avoid misjudgments caused by data fluctuations. Therefore, data with total scores near the threshold (e.g., ±5% of the threshold) are added to the manual review list to make them more accurate.

[0113] Feedback learning: Expert judgments are fed back to the system as labels, allowing for fine-tuning of rule weights or training of more advanced classification models, thus enabling continuous evolution of analytical capabilities. The specific implementation is as follows: ① Construction of labeled samples: Normal days confirmed by experts are marked as negative samples and abnormal days are marked as positive samples. Each sample carries multi-dimensional features such as total fluctuation deviation, DTW distance, number of extreme points, proportion of change rate exceeding limit, date type, season, user type, etc., to build an incremental labeled sample library; ② Rule weight fine-tuning: Every 50 to 100 labeled samples are accumulated, incremental logistic regression or incremental perceptron algorithm is used, with each rule output as feature and expert label as supervision target, to minimize cross-entropy loss and adaptively update the weights of each judgment dimension to reduce misjudgment and missed judgment. ③ Advanced classification model training: When the number of labeled samples is ≥300, use multi-dimensional judgment indicators as input and expert judgment labels as output to train high-generalization binary classification models such as random forest and lightweight gradient boosting tree (GBRT) to output anomaly probabilities; the model is divided into training set and validation set in a 7:3 ratio, and early stopping is used to prevent overfitting; after training, it is used in parallel with the rule engine for voting decision to improve judgment accuracy; ④ Iterative Closed-Loop Results: After optimization, metrics such as false positive rate, false negative rate, amount of manual review, and anomaly recall rate are statistically analyzed. When the amount of manual review decreases by ≥20% and the judgment accuracy increases by ≥5%, the system is officially updated and launched, forming a closed-loop evolution mechanism of "judgment-decision-optimization-deployment". Step 5: As Figure 5 As shown, when a day is identified as an abnormal fluctuation day, a second dynamic check is performed on the approved baseline, including model performance drift detection and baseline backtracking consistency analysis. Based on the results of the second dynamic check, the CA-DTN model is optimized for parameters, and the optimized CA-DTN model is used to output the final approved baseline load curve.

[0114] Step 5.1, Model performance drift detection.

[0115] Using the actual load data from the most recent assessment period (e.g., the past 30 days) as a new test set, the key forecast accuracy metrics of the current online CA-DTN model are recalculated, including mean absolute percentage error (MAPE) and root mean square error (RMSE).

[0116] The statistical process control (SPC) method was used to plot the daily MAPE values ​​on a control chart. The center line was set as the mean MAPE value of the historical normal period (such as the first 90 days after the initial deployment), and the upper control limit was the mean plus 3 times the standard deviation.

[0117] If the MAPE value is higher than the upper control limit for 7 consecutive days, or if the MAPE value shows a statistically significant monotonically increasing trend (through the Mann-Kendall trend test, significance level α=0.05), then the model is considered to have experienced performance drift.

[0118] Once drift is detected, model retraining is automatically triggered.

[0119] Step 5.2, baseline backtracking consistency analysis.

[0120] Select a set of dates for which complete actual load data has been obtained historically and for which assessments have been completed previously (e.g., all response event days within the past 90 days).

[0121] Using the latest CA-DTN model, the original baseline is re-predicted for each of the above dates, and the dynamic correction in step 3 is performed to obtain the newly calculated approved baseline.

[0122] Will The approved baseline used in the original assessment on that date By comparing each point, the following three quantitative indicators were calculated: ① The longest duration (number of points) during which the relative error at each point continuously exceeds 5%; ②The daily average relative error for the whole day: ; ③ Average absolute error during the response period (take the absolute value, do not divide by the load value).

[0123] A systematic difference is determined to exist if any of the following conditions are met: ① The number of time points where the relative error per point continuously exceeds 5% is ≥ 30; ②The daily average relative error is >3%; ③ Baseline rated capacity with an average absolute error > 4% during the response period.

[0124] If three or more consecutive dates are identified as systematic discrepancies, model retraining will be automatically triggered, and a discrepancy report will be generated, recording the affected historical evaluation periods.

[0125] Step 5.3: After determining whether there is drift or systematic difference, the model retraining is automatically triggered. The process is as follows: (1) Training data preparation: Extract all “normal days” (non-abnormal fluctuation days) samples confirmed in step 4 from historical data, take the data of the most recent 180 days in chronological order, and construct the time series feature sequence and context feature vector according to the method in step 1.

[0126] (2) Training set / validation set partitioning: The above samples are divided into training set, validation set and test set in a ratio of 7:1.5:1.5 (keeping the time order and not using future data for training).

[0127] (3) Model retraining: Keep the CA-DTN model architecture exactly the same as in step 2, and retrain on the training set using the same composite loss function (MSE loss + smoothness loss + controllability guidance loss) and optimizer (AdamW, initial learning rate 0.001) as in the initial training. Use early stopping (patience value = 10) to stop training on the validation set to prevent overfitting.

[0128] (4) Performance verification: Evaluate the MAPE and RMSE metrics of the retrained CA-DTN model on the test set. If the MAPE on the test set decreases by more than 0.5 percentage points compared to the model before drift detection is triggered, the online model will be automatically replaced; otherwise, the original model will be retained and an alarm will be recorded for manual review.

[0129] (5) Model deployment: New model parameters are written into the production environment through version management for use in all subsequent prediction tasks.

[0130] Step 5.4, Final Approval Baseline Output: Steps 2 and 3 are re-executed using the retrained and optimized CA-DTN model to generate a new approved baseline as the final approved baseline. .

[0131] If model retraining is not triggered, the approved baseline output in step 3 will be used directly as the final approved baseline. .

[0132] The final approved baseline is output to step 6 for response effect evaluation.

[0133] Step 6: Response Effectiveness Evaluation. The actual load curve is compared with the final approved baseline load curve verified in Steps 3-5 during the response period to accurately calculate key performance indicators such as response volume, response completion rate, and regulation accuracy. The system automatically generates a structured evaluation report, presenting the entire process and results in the form of visual charts and data analysis, supporting auditing, settlement, and appeal review.

[0134] Core indicator calculation: Actual response volume: ; in, In response to the start time, For the response end time, To finally verify the baseline load, This is the actual load curve.

[0135] Response completion rate: ; Where Target is the target response quantity.

[0136] Adjustment accuracy:

[0137] in, Actual load curve, This is the target load curve.

[0138] The assessment report is presented in the form of visual charts and data analysis, including a baseline load curve comparison chart, a statistical table of response performance indicators, and an explanation of abnormal situations.

[0139] The following two specific embodiments, together with the accompanying drawings, further illustrate the implementation process and technical effects of the present invention in detail.

[0140] Example 2 Day-ahead baseline assessment and response effectiveness calculation for the spot market This embodiment uses the participation of a virtual power plant in day-ahead electricity spot market transactions as a case study to demonstrate the complete application of the method of the present invention in long-term, high-precision baseline forecasting and fair assessment.

[0141] 1. Scenario and Data Preparation Evaluation subject: A virtual power plant that aggregates the controllable loads of 20 manufacturing enterprises in an industrial park.

[0142] Market type: Electricity spot market, requiring day-ahead bidding, and forecasting baseline load for 96 time points (15-minute intervals) from 00:00 to 24:00 the following day.

[0143] Data source: Historical load data: Aggregated active power data of the virtual power plant over the past two years, in 15-minute granularity.

[0144] Meteorological data: Temperature, humidity, and wind speed data from local weather stations during the same period.

[0145] Market and calendar data: historical electricity prices, holiday information, weekday types.

[0146] Data preprocessing: For obvious metering errors and missing data caused by communication interruptions in historical load data, linear interpolation between the preceding and following time periods, combined with daily data from the same period, is used to compensate for the deficiencies.

[0147] The Isolation Forest algorithm is used to detect and remove abnormal daily load curves (such as days when the entire plant is shut down) from historical data.

[0148] All numerical features (load, temperature, etc.) are subjected to maximum-min normalization and mapped to the [0,1] interval.

[0149] 2. CA-DTN Model Training and Prediction Feature construction: Temporal feature sequences ( The load and temperature sequences for 10 consecutive days (960 time points in total) prior to the forecast date are selected as input. Cyclic coding features such as "hour" and "day type" are also added.

[0150] Contextual feature vector (C): market_type=[0,1,0] (representing the spot market), target_day_type=weekday, season=summer, is_holiday=no.

[0151] Model training: The dataset is divided into a training set (70%), a validation set (15%), and a test set (15%).

[0152] The AdamW optimizer was used with an initial learning rate of 0.001, along with cosine annealing scheduling.

[0153] Use a composite loss function and set weights. (Ensure the accuracy of basic predictions) (Balancing smoothness and prediction accuracy) (Enhanced guidance based on optimal controllable baseline). Among them... The optimization simulation was conducted using historical data with the goal of minimizing intraday load fluctuations.

[0154] Set the batch size to 64 and use early stopping (patience value = 20) to prevent overfitting.

[0155] Baseline prediction: The constructed predicted daily features are input into the trained CA-DTN model to obtain the original baseline load curve. A total of 96 prediction points.

[0156] Performance Results: On the independent test set, the model predicted a MAPE of 2.8% and a coefficient of determination of [missing information]. The accuracy was 0.945, and the average prediction error for peak load was 125.6 kW, both significantly better than the compared LSTM model (MAPE 3.5%). 0.921).

[0157] 3. Dynamic adjustment of the response event date (assuming a demand response event occurs the next day). Event Information: Assume that the power grid issues a demand response event between 14:00 and 16:00 on the forecast date.

[0158] Obtain actual data: After the event ends, obtain the actual load curve for the entire day. .

[0159] Execute the correction algorithm (see Figure 3 process): Phase 1: Selecting Non-Response Periods The period is defined as 2 hours prior to the start of the event (12:00-14:00). This time period is calculated. (less than the threshold) =5%), therefore no global scaling adjustment is performed. .

[0160] Phase Three: Calculation Time residual The correction function for the entire day was fitted using cubic spline interpolation. The final corrected baseline was obtained. After correction, the fitting error during non-response periods is almost zero, and the baseline during response periods is smoothly adjusted.

[0161] 4. Verification and final confirmation Load fluctuation analysis: The system automatically analyzes the actual load curve for the day. Indicators such as total load fluctuation and DTW morphological distance do not trigger any abnormal rules, thus classifying it as a normal day.

[0162] Secondary verification (periodic trigger): During this week's routine verification, the system did not detect any performance drift in the CA-DTN model, and the consistency of historical baseline backtracking was good.

[0163] Expert review: This case did not trigger manual review. The system automatically confirmed. This serves as the final approved baseline.

[0164] 5. Response effect calculation and reporting Calculate core metrics: .

[0165] If the response target is 100MWh, then the response completion rate = 105.2 / 100*100% = 105.2%.

[0166] The 2-norm distance (Euclidean distance) between the actual load curve and the target curve was calculated, yielding an adjustment accuracy of 96.3%.

[0167] Report generation: The system automatically generates PDF and web visualization reports, including a three-curve comparison chart of the forecast baseline, the corrected baseline, and the actual load, detailed data tables, and complete indicator calculation results and correction logs.

[0168] Example 3 Real-time baseline assessment and anomaly handling for the ancillary services market This embodiment demonstrates the application of the present invention in an auxiliary service market (such as frequency modulation) that requires high real-time performance and high frequency, and demonstrates the abnormal handling process when an unreported maintenance occurs.

[0169] 1. Scenarios and Challenges Market type: Ancillary services market (Automatic Generation Control (AGC) / Frequency Regulation).

[0170] The core challenge is the need to predict the baseline for the next few minutes in real-time (e.g., every 4 seconds) or near real-time (every minute) to calculate the frequency modulation contribution at the second or minute level. This places extremely high demands on the model's computational speed and ability to capture short-term fluctuations.

[0171] Unexpected situation: On a certain frequency regulation service day, a large piece of equipment in the virtual power plant unexpectedly failed and was shut down for maintenance (without prior notification), resulting in a sudden drop in load.

[0172] 2. Short-term forecasting applications of the CA-DTN model Model adaptation: Use the same CA-DTN model architecture, but adjust the input and output.

[0173] enter : Ultra-short-term load and frequency deviation data for the past hour (one point per minute, 60 points in total).

[0174] Context C: market_type=[1,0,0] (auxiliary service market), real-time system frequency status.

[0175] Real-time prediction: The model is deployed on a high-performance computing server and can continuously predict the baseline load value for the next 5 minutes with a latency of up to one second. .

[0176] 3. Occurrence of anomalies and system response Equipment failure: At 15:30 in the afternoon, a critical production equipment unexpectedly stopped.

[0177] Load fluctuation analysis engine provides real-time alarms: The abnormal rate of change rule is triggered first: the load decreases by more than 5 standard deviations of the historical normal minute-level rate of change within 1 minute.

[0178] The total load fluctuation rule was subsequently triggered: the average load over the next 30 minutes was more than 3σ lower than the historical average for the same period.

[0179] Based on comprehensive judgment, if the total weighted voting score exceeds the threshold, the system will mark the period as "abnormal fluctuation" and issue an alarm in real time.

[0180] Triggering secondary verification and manual intervention: Because it is a real-time scenario with drastic fluctuations, the system automatically triggers an emergency verification process and pushes relevant data and judgment results to the expert review workbench of the operations staff.

[0181] The on-duty expert viewed the real-time curves, historical comparisons, and rule trigger details through the workbench, and confirmed with the power plant by phone that it was an unreported emergency maintenance.

[0182] The expert performed the "confirm anomaly and activate the backup baseline plan" operation on the workbench.

[0183] 4. Anomaly Handling and Baseline Recalculation Enable backup baseline: According to preset rules, when a certain day is marked as abnormal and confirmed by manual means, the system automatically uses the load data of several recent normal days of the same type of virtual power plant to generate an alternative baseline for the abnormal period through the moving average method.

[0184] Performance Evaluation: An alternative baseline was used to calculate the frequency regulation performance of the virtual power plant during the fault (not its responsibility), avoiding misjudgments of negative contributions due to its own reasons and ensuring the fairness of the evaluation. Simultaneously, the event was recorded in the case library for future optimization of the judgment rules.

[0185] 5. Advantages highlighted in this embodiment Real-time performance: The CA-DTN model and analysis engine can meet the real-time processing requirements at the second to minute level.

[0186] Strong robustness: In the face of sudden and serious anomalies, multiple safeguard mechanisms (real-time assessment, emergency verification, manual confirmation, and backup plans) are quickly activated to ensure the continuity and impartiality of the assessment system.

[0187] Closed-loop learning: The abnormal event and its handling results can be used as feedback data to optimize the parameters of the load fluctuation judgment model and improve the accuracy of identifying similar emergency maintenance in the future.

[0188] The beneficial effects of this invention are as follows: First, compared with existing technologies, by introducing a prediction model that integrates contextual attention and deep time-series networks (CA-DTN), the prediction accuracy and scenario adaptability of baseline load are significantly improved. It can achieve high-precision prediction with an average absolute percentage error of less than 3% in diverse electricity market environments, effectively supporting fair accounting of response quantities. Second, by constructing a multi-layered guarantee mechanism including dynamic baseline correction, maintenance day reporting, intelligent load fluctuation analysis, and secondary verification, the system possesses strong anti-interference capabilities and self-optimization capabilities, providing reliable guarantees across the entire chain from data source to evaluation results, greatly improving the fairness and traceability of evaluation results. Furthermore, this method considers engineering practicality; the model and mechanism design support efficient parallel computing, meeting the timeliness requirements of day-ahead and real-time evaluations, and can be seamlessly integrated into existing electricity market operation platforms. Finally, by introducing "controllability guidance" into the training objective, the baseline load not only reflects historical patterns but also closely approximates the optimal and stable dispatchable state of the virtual power plant, thereby guiding resource aggregators to improve operational levels and enhancing the overall competitiveness and value creation capabilities of the virtual power plant in the electricity market.

[0189] Example 4 This invention also discloses a virtual power plant response performance evaluation system based on the aforementioned virtual power plant response performance evaluation method, comprising: The data and feature extraction module collects historical aggregated power data, meteorological data, calendar information, and market transaction records of the target virtual power plant, constructs a time-series feature sequence and a context feature vector, and performs preprocessing; the time-series feature sequence includes a fluctuating continuous time series, and the context feature vector includes static information of the operating scenario features; Baseline prediction module: Input the preprocessed temporal feature sequence and context feature vector into the constructed deep temporal network CA-DTN model to predict the original baseline load curve for the response event day; Dynamic correction module: acquires actual load data and event information for the response event day, dynamically corrects the original baseline load curve, and generates an approved baseline; Fluctuation Day Judgment Module: Based on a preset set of multi-dimensional load fluctuation rationality judgment rules, the module judges the approved baseline to determine whether it belongs to an abnormal fluctuation day. Secondary dynamic verification and retraining module: When an abnormal fluctuation day is identified, a secondary dynamic verification is performed on the approved baseline, including model performance drift detection and baseline backtracking consistency analysis; based on the results of the secondary dynamic verification, the CA-DTN model is retrained, and the final approved baseline load curve is output using the optimized CA-DTN model; Effect evaluation module: Compares the actual load curve of the response event day with the final approved baseline load curve, calculates the actual response amount, response completion rate and regulation accuracy index of the virtual power plant, and generates a visual evaluation report that includes curve comparison, index statistics and anomaly description.

[0190] Also includes: Data interaction interface module: Supports standardized data interaction with multiple source systems such as the power market trading platform, virtual power plant user application terminal, and meteorological service platform, realizing the automated flow of data access, command issuance, and result feedback, and ensuring data security and integrity.

[0191] The assessment report generation module automatically integrates curve comparisons, indicator statistics, anomaly explanations, and historical trend analysis to generate assessment reports in multiple formats, including PDF and web visualization, meeting the needs of various scenarios such as settlement audits, appeal reviews, and regulatory supervision.

[0192] Example 5 The present invention also proposes a terminal, including a processor and a storage medium: The storage medium is used to store instructions; The processor is used to perform the steps of the above method according to the instructions.

[0193] Example 6 The present invention also proposes a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the above-described method.

[0194] This disclosure can be a system, method, and / or computer program product. A computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for causing a processor to implement various aspects of this disclosure.

[0195] Computer-readable storage media can be tangible devices capable of holding and storing instructions for use by an instruction execution device. Computer-readable storage media can be, for example—but not limited to—electrical storage devices, magnetic storage devices, optical storage devices, electromagnetic storage devices, semiconductor storage devices, or any suitable combination of the foregoing. More specific examples (a non-exhaustive list) of computer-readable storage media include: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), static random access memory (SRAM), portable compact disc read-only memory (CD-ROM), digital multifunction disc (DVD), memory sticks, floppy disks, mechanical encoding devices, such as punch cards or recessed protrusions storing instructions thereon, and any suitable combination of the foregoing. The computer-readable storage media used herein are not to be construed as transient signals themselves, such as radio waves or other freely propagating electromagnetic waves, electromagnetic waves propagating through waveguides or other transmission media (e.g., light pulses through fiber optic cables), or electrical signals transmitted through wires.

[0196] The computer-readable program instructions described herein can be downloaded from computer-readable storage media to various computing / processing devices, or downloaded via a network, such as the Internet, local area network, wide area network, and / or wireless network, to an external computer or external storage device. The network may include copper transmission cables, fiber optic transmission, wireless transmission, routers, firewalls, switches, gateway computers, and / or edge servers. A network adapter card or network interface in each computing / processing device receives the computer-readable program instructions from the network and forwards them to the computer-readable storage media in the respective computing / processing device.

[0197] Computer program instructions used to perform the operations of this disclosure may be assembly instructions, instruction set architecture (ISA) instructions, machine instructions, machine-dependent instructions, microcode, firmware instructions, status setting data, or source code or object code written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Smalltalk, C++, etc., and conventional procedural programming languages ​​such as the "C" language or similar programming languages. The computer-readable program instructions may execute entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving a remote computer, the remote computer may be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or may be connected to an external computer (e.g., via the Internet using an Internet service provider). In some embodiments, electronic circuitry, such as programmable logic circuitry, field-programmable gate arrays (FPGAs), or programmable logic arrays (PLAs), is personalized by utilizing the status information of the computer-readable program instructions to implement various aspects of this disclosure.

[0198] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the protection scope of the claims of the present invention.

Claims

1. A method for evaluating the response effect of a virtual power plant, characterized in that, include: Step 1: Collect historical aggregated power data, meteorological data, calendar information, and market transaction records of the target virtual power plant, construct a time-series feature sequence and a context feature vector, and perform preprocessing; the time-series feature sequence includes a fluctuating continuous time series, and the context feature vector includes static information of the operating scenario features; Step 2: Input the preprocessed temporal feature sequence and contextual feature vector into the constructed deep temporal network CA-DTN model to predict the original baseline load curve for the response event day; Step 3: Obtain the actual load data and event information for the response event day, dynamically correct the original baseline load curve, and generate the approved baseline; Step 4: Based on the preset multi-dimensional load fluctuation rationality judgment rule set, the approved baseline is judged to determine whether it belongs to an abnormal fluctuation day; Step 5: When an abnormal fluctuation day is identified, a second dynamic verification is performed on the approved baseline, including model performance drift detection and baseline backtracking consistency analysis; based on the results of the second dynamic verification, the CA-DTN model is retrained, and the final approved baseline load curve is output using the optimized CA-DTN model. Step 6: Compare the actual load curve of the response event day with the final approved baseline load curve, calculate the actual response amount, response completion rate and regulation accuracy index of the virtual power plant, and generate a visual evaluation report that includes curve comparison, index statistics and anomaly description.

2. The method for evaluating the response effect of a virtual power plant according to claim 1, characterized in that: In step 2, the CA-DTN model includes an input layer, a context branch, a temporal branch, and an output layer; The input layer receives temporal feature sequences and context feature vectors in parallel. The context branch includes an embedding layer and a context attention module, which ultimately outputs a weighted context vector. The temporal branch includes a vector augmentation layer, a deep temporal encoder, and a temporal attention module, which ultimately outputs a weighted temporal vector. The output layer concatenates the weighted context vector and the weighted time series vector and then obtains the baseline load curve through a fully connected structure.

3. The method for evaluating the response effect of a virtual power plant according to claim 2, characterized in that: In step 2, the context branch includes the embedding layer and the context attention module: In the embedding layer, categorical features in the context feature vector are embedded and encoded, while numerical features are standardized to form a dense vector. In the context attention module, The features in the dataset are divided into at least two feature groups based on their categories. Contextual attention weights are calculated for each feature group using independent multilayer perceptron and softmax layers. Based on the contextual attention weights of each feature group, the... The corresponding feature subvectors are weighted, and then the weighted features in all feature groups are summed to generate a weighted context vector.

4. The method for evaluating the response effect of a virtual power plant according to claim 2, characterized in that: In step 2, the temporal branch includes a vector augmentation layer, a deep temporal encoder, and a temporal attention module. The specific implementation method is as follows: In the vector enhancement layer, the CA-DTN model also includes a deep temporal network module, which receives temporal feature sequences. With the context representation vector The specific processing procedure is as follows: Temporal feature embedding: embedding temporal feature sequences Layer mapping to a high-dimensional space yields temporal feature embeddings. The weighted context vector output from the context branch is broadcast to each time step and concatenated with the temporal feature embedding to form an enhanced temporal input; The basic architecture of a deep temporal encoder is either BiLSTM or GRU, which encodes the enhanced temporal input: a multi-layer bidirectional LSTM is used to process the enhanced temporal input, obtain the forward and backward hidden states, and concatenate the states at the corresponding time steps to obtain the temporal representation. ; In the temporal attention module, a similarity score is calculated between the trainable query vector and each hidden state in H, and the temporal attention weights are obtained by normalizing the similarity scores. The weighted sum of H is then calculated based on the temporal attention weights to generate a weighted temporal vector.

5. The method for evaluating the response effect of a virtual power plant according to claim 1, characterized in that: In step 3, the dynamic correction method is a three-stage correction method, specifically: The first phase involves selecting non-execution periods on the response event day that are not affected by market orders. Calculate the actual load during this period. Average relative deviation of baseline load curve ;like Less than the preset threshold Then, the third stage of residual fitting is entered to generate the final baseline; Otherwise, after performing the second stage of preliminary offset correction, the corrected baseline load curve is then input into the third stage. The second stage involves fitting and correcting the original baseline load curve: During the execution period of the response event Using average relative offset The original baseline load curve is weighted to obtain the corrected baseline load curve; The third stage involves calculating non-execution periods. Within this, the actual load is compared with the original baseline load curve or the corrected baseline load curve obtained after the second stage. sequence ; Spline interpolation or Gaussian process regression methods are used to analyze discrete residual points. By fitting the data, a smoothing correction function defined over the entire time period is obtained. Will The approved baseline is generated by adding the original baseline load curve or the modified baseline load curve obtained in the second stage.

6. The method for evaluating the response effect of a virtual power plant according to claim 1, characterized in that: In step 4, the multi-dimensional load fluctuation rationality assessment rule set includes judgment rules for five dimensions: total load fluctuation, abnormal curve shape, peak / trough detection, and abnormal rate of change. Total fluctuation dimension: Calculate the total electricity consumption on the target day. If its daily average value for the same period in history is The absolute difference exceeds 100 times the historical standard deviation If so, an exception flag is triggered, where These are preset coefficients; Anomaly dimension of curve shape: Calculate the target daily load curve L and the historical typical daily load curve. Dynamic time warping distance If this distance exceeds the historical normal daily DTW distance distribution If the quantile is reached, an exception flag is triggered; Peak / trough detection dimension: Employs a sliding window statistical method to detect peaks / troughs that consistently exceed the mean within the window. If the number of local extreme points exceeds a threshold, or if the isolated forest algorithm determines that the daily curve is an overall outlier, an anomaly marker is triggered. Anomaly dimension of rate of change: Calculate the point-by-point difference ΔL(t) of the load, and count whether its absolute value exceeds the historical normal difference range. The proportion, if the proportion exceeds a preset threshold If so, an exception flag will be triggered.

7. The method for evaluating the response effect of a virtual power plant according to claim 1, characterized in that: In step 4, the specific method for determining whether a day belongs to abnormal fluctuations is as follows: Based on the rule set for judging the rationality of load fluctuations in a multi-dimensional manner, the judgment result of each rule is generated for the approved baseline. The judgment results are then weighted and voted on to obtain a weighted total score for anomalies. If the weighted abnormal total score exceeds the set total score threshold, it is determined to be an abnormal fluctuation day.

8. The method for evaluating the response effect of a virtual power plant according to claim 1, characterized in that: Step 5 involves secondary dynamic verification, including model performance drift detection and baseline backtracking consistency analysis. The implementation methods include: Model performance drift detection uses the latest historical load data for a set period as a new test set to recalculate the prediction accuracy indicators of the current CA-DTN model, including the mean absolute percentage error (MAPE). Statistical process control methods are used, with the center line set as the mean of the mean absolute percentage error over a set historical normal period, and the upper control limit set as the mean plus three standard deviations. If the mean absolute percentage error for a consecutive set number of days is higher than the upper control limit, or if the mean absolute percentage error shows a statistically significant monotonically increasing trend, then the model is determined to have experienced performance drift. Baseline retrospective consistency analysis selects a set of dates for which complete historical actual load data has been obtained and previously assessed. Using the latest CA-DTN model, the original baseline is re-predicted for each of these dates, and the dynamic correction in step 3 is performed to obtain the newly calculated approved baseline. ;Will The approved baseline used in the original assessment on that date A point-by-point comparison is performed, and quantitative indicators are calculated, including the longest duration during which the point-by-point relative error continuously exceeds the set error threshold, the daily average relative error for the whole day, and the average absolute error within the response period. If the longest duration is not less than the set time threshold, or the daily average relative error for the whole day is greater than the set threshold, or the average absolute error within the response period is greater than the set threshold of the baseline rated capacity, then a systematic difference is determined to have occurred, triggering the model retraining of the CA-DTN model.

9. The method for evaluating the response effect of a virtual power plant according to claim 1, characterized in that: In step 5, the CA-DTN model is optimized based on the results of the secondary dynamic verification. The optimized CA-DTN model is then used to output the final approved baseline load curve. The specific steps include: Extract all normal day samples confirmed in step 4 from historical data, take the data of the most recent set number of days in chronological order, and construct the time series feature sequence and context feature vector according to the method in step 1. The above samples are divided into training set, validation set and test set; Keep the CA-DTN model architecture exactly the same as in step 2, and retrain it on the training set; Evaluate the MAPE and RMSE metrics of the retrained CA-DTN model on the test set. If the MAPE on the test set decreases by more than a set threshold compared to the model before drift detection is triggered, the model is automatically replaced; otherwise, the original model is retained and an alarm is recorded.

10. A virtual power plant response performance evaluation system utilizing the method described in any one of claims 1-9, comprising a data and feature extraction module, a baseline prediction module, a dynamic correction module, a fluctuation day judgment module, a secondary dynamic verification and retraining module, and a performance evaluation module, characterized in that: The data and feature extraction module collects historical aggregated power data, meteorological data, calendar information, and market transaction records of the target virtual power plant, constructs a time-series feature sequence and a context feature vector, and performs preprocessing; the time-series feature sequence includes a fluctuating continuous time series, and the context feature vector includes static information of the operating scenario features; The baseline prediction module inputs the preprocessed temporal feature sequence and context feature vector into the constructed deep temporal network CA-DTN model to predict the original baseline load curve for the response event day. The dynamic correction module acquires the actual load data and event information for the response event day, dynamically corrects the original baseline load curve, and generates an approved baseline. The fluctuation day judgment module, based on a preset multi-dimensional load fluctuation rationality judgment rule set, judges the approved baseline to determine whether it belongs to an abnormal fluctuation day; The secondary dynamic verification and retraining module performs a secondary dynamic verification of the approved baseline when an abnormal fluctuation day is identified. This includes model performance drift detection and baseline backtracking consistency analysis. Based on the results of the secondary dynamic verification, the CA-DTN model is retrained, and the optimized CA-DTN model is used to output the final approved baseline load curve. The effect evaluation module compares the actual load curve of the response event day with the final approved baseline load curve, calculates the actual response volume, response completion rate and regulation accuracy of the virtual power plant, and generates a visual evaluation report that includes curve comparison, indicator statistics and anomaly descriptions.

11. A terminal, comprising a processor and a storage medium; characterized in that: The storage medium is used to store instructions; The processor is configured to operate according to the instructions to perform the steps of the method according to any one of claims 1-9.

12. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the program implements the steps of the method according to any one of claims 1-9.

Images