An adaptive integrated electricity price forecasting method based on failure scenario detection
Patent Information
- Application Number
- CN202610890371.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-06-18
- Publication Date
- 2026-09-11
AI Technical Summary
现有方法未能显式识别这些失效场景，更无法根据场景动态调整预测策略，导致在关键时段的预测精度无法保障，因此，亟需一种能够感知市场运行状态、自适应组合预测模型的基于失效场景检测的自适应集成电力价格预测方法
[0027]Compared with existing technologies, the beneficial effects of this invention are as follows: The adaptive integrated electricity price forecasting method based on failure scenario detection of this invention avoids model failure under complex operating conditions by explicitly identifying failure scenarios and calling dedicated models. The prediction accuracy is improved by an average of 15%-25%, improving prediction robustness, providing confidence and risk warnings, assisting in trading decisions, reducing decision risks, enhancing decision reliability, and the online update mechanism enables the system to dynamically adapt to changes in market rules and the emergence of new scenarios. It has strong adaptability, can flexibly expand to new failure scenarios and dedicated models, and is easy to maintain and upgrade.
Smart Images
Figure CN122736671A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of electricity price forecasting technology, specifically an adaptive integrated electricity price forecasting method based on failure scenario detection. Background Technology
[0002] With the deepening of power market reform, electricity price forecasting has become the core of market participants' trading strategies. Existing forecasting methods mostly employ single models (such as XGBoost and LSTM) or simple ensembles (such as averaging and weighted averages). However, electricity prices are affected by multiple factors, resulting in various "failure scenarios"—in which the predictive performance of conventional models drops sharply, and even directional errors occur. Typical failure scenarios include: blocking failure, ramp-up failure, nonlinear interval failure, and uncertainty amplification failure. Existing methods fail to explicitly identify these failure scenarios, and cannot dynamically adjust forecasting strategies according to the scenarios, leading to a lack of accuracy in forecasting during critical periods. Therefore, there is an urgent need for an adaptive ensemble electricity price forecasting method based on failure scenario detection that can sense the market operating status and adaptively combine forecasting models. Summary of the Invention
[0003] This invention provides an adaptive integrated electricity price forecasting method based on failure scenario detection to address the problems mentioned in the background section.
[0004] According to an embodiment of the present invention, an adaptive integrated electricity price forecasting method based on failure scenario detection is provided, comprising the following steps:
[0005] Step S1: Historical data collection and preprocessing, collecting historical electricity market data, calculating the bidding space, cleaning and normalizing the data, and dividing it into training set, validation set, and test set;
[0006] Step S2: Failure scenario definition and detection model training. Define at least one failure scenario, construct sample labels based on historical data, extract features, train a scenario classifier and use it to detect the probability of belonging to each failure scenario in real time.
[0007] Step S3: Training the general model and the special model. Use all historical data to train the baseline prediction model, and for each failure scenario, select a subset of historical data for that scenario to train a special prediction model.
[0008] Step S4: Real-time failure scenario detection. During online prediction, the current real-time data is obtained and input into each scenario classifier to obtain the probability of belonging to each failure scenario in the current time period.
[0009] Step S5: Adaptive ensemble prediction. The ensemble weights of each model are dynamically calculated based on the scenario probability and the preset model confidence coefficient. The current features are input into each model to obtain the predicted values of each model. The predicted values are then weighted and fused to obtain the final electricity price prediction. The prediction confidence is calculated based on the scenario probability and the consistency of the model prediction. If the confidence is lower than the confidence threshold or the probability of any failure scenario is higher than the risk threshold, an early warning is generated.
[0010] Step S6: Online model update. New data generated during the online prediction phase is collected periodically. The performance of the general model and each special model is evaluated. When the performance of any model drops below a preset threshold or a new failure scenario is detected, retraining is triggered, and the original model is replaced with the updated model.
[0011] As a further aspect of the present invention: In step S1, the historical electricity market data includes at least one of the following: central dispatch load, new energy output, external power transmission, cross-sectional power flow, unit operating capacity, ramp-up capability, fuel price, and temperature. The bidding space is calculated as: central dispatch load - new energy output - external power transmission.
[0012] As a further aspect of the present invention: In step S2, the training scene classifier uses a random forest or XGBoost model, and the probability of belonging to each failure scene in the current time period is... Where i represents the scenario type, the extracted features include at least one of the following: cross-sectional load rate, load change rate, bidding spatial quantile, and prediction error, and the failure scenario includes at least one of the following four:
[0013] In a congested scenario, the load rate of the critical section exceeds the preset high load rate threshold and the duration exceeds the duration threshold.
[0014] In the ramp-up scenario, the net load change rate exceeds the preset percentage threshold of the system ramp-up capacity, and the occurrence occurs within the preset peak period.
[0015] In non-linear interval scenarios, the bidding space is either lower than the historical low quantile threshold or higher than the high quantile threshold.
[0016] In scenarios where uncertainty is amplified, the absolute value of the prediction error for new energy sources exceeds the high quantile threshold of the historical distribution.
[0017] As a further aspect of the present invention: in step S3, the general model and the special model adopt the same or different benchmark prediction models, and the benchmark prediction model includes AutoGluon TabularPredictor.
[0018] As a further aspect of the present invention: in step S4, the current real-time data includes at least one of the following: cross-sectional load rate, net load change rate, bidding spatial quantile, new energy prediction deviation, and time / season. Each scenario classifier includes a congestion scenario classifier, a ramp scenario classifier, a nonlinear interval classifier, and an uncertainty amplification classifier.
[0019] As a further aspect of the present invention: in step S5, the formula for calculating the integration weight is:
[0020] ;
[0021] in, These correspond to four different failure scenarios. Corresponding to the general model, The confidence coefficient of the model. The prediction accuracy of each model on the validation set can be preset, or dynamically adjusted through online learning. With no probability of failure and Or normalization can be used. These represent the probabilities of four failure scenarios.
[0022] As a further aspect of the present invention: in step S5, the predicted values of each model include: the predicted value of the general model. Blocking-specific model predictions Predicted values from a hill-climbing model Nonlinear dedicated model prediction value and uncertainty-specific model predictions The weighted fusion calculation formula is as follows:
[0023] ;
[0024] in, For the final electricity price forecast, These are the predicted values for each model.
[0025] As a further aspect of the present invention: in step S5, the integrated weights can also be learned by a meta-learner based on historical data, wherein the meta-learner includes linear regression or a neural network.
[0026] As a further aspect of the present invention: in step S6, the retraining adopts a sliding window approach, using only the updated data from the most recent period, or adopts an incremental learning approach to update the model parameters and retrain the corresponding scene classifier.
[0027] Compared with existing technologies, the beneficial effects of this invention are as follows: The adaptive integrated electricity price forecasting method based on failure scenario detection of this invention avoids model failure under complex operating conditions by explicitly identifying failure scenarios and calling dedicated models. The prediction accuracy is improved by an average of 15%-25%, improving prediction robustness, providing confidence and risk warnings, assisting in trading decisions, reducing decision risks, enhancing decision reliability, and the online update mechanism enables the system to dynamically adapt to changes in market rules and the emergence of new scenarios. It has strong adaptability, can flexibly expand to new failure scenarios and dedicated models, and is easy to maintain and upgrade. Attached Figure Description
[0028] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments of the present invention will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0029] Figure 1 This is a first flowchart of an adaptive integrated electricity price prediction method based on failure scenario detection, provided in an embodiment of the present invention.
[0030] Figure 2 This is a second flowchart of an adaptive integrated electricity price prediction method based on failure scenario detection, provided in an embodiment of the present invention.
[0031] Figure 3 This is a structural diagram of the failure scenario definition and detection model training provided in the embodiments of the present invention.
[0032] Figure 4 A flowchart of adaptive ensemble prediction provided for an embodiment of the present invention. Detailed Implementation
[0033] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0034] Please see Figure 1 An adaptive integrated electricity price forecasting method based on failure scenario detection includes the following steps:
[0035] Step S1: Historical data collection and preprocessing, collecting historical electricity market data, calculating the bidding space, cleaning and normalizing the data, and dividing it into training set, validation set, and test set;
[0036] Step S2: Failure scenario definition and detection model training. Define at least one failure scenario, construct sample labels based on historical data, extract features, train a scenario classifier and use it to detect the probability of belonging to each failure scenario in real time.
[0037] Step S3: Training the general model and the special model. Use all historical data to train the baseline prediction model, and for each failure scenario, select a subset of historical data for that scenario to train a special prediction model.
[0038] Step S4: Real-time failure scenario detection. During online prediction, the current real-time data is obtained and input into each scenario classifier to obtain the probability of belonging to each failure scenario in the current time period.
[0039] Step S5: Adaptive ensemble prediction. The ensemble weights of each model are dynamically calculated based on the scenario probability and the preset model confidence coefficient. The current features are input into each model to obtain the predicted values of each model. The predicted values are then weighted and fused to obtain the final electricity price prediction. The prediction confidence is calculated based on the scenario probability and the consistency of the model prediction. If the confidence is lower than the confidence threshold or the probability of any failure scenario is higher than the risk threshold, an early warning is generated.
[0040] Step S6: Online model update. New data generated during the online prediction phase is collected periodically. The performance of the general model and each special model is evaluated. When the performance of any model drops below a preset threshold or a new failure scenario is detected, retraining is triggered, and the original model is replaced with the updated model.
[0041] In another embodiment, Figure 2 This is a second flowchart of an adaptive integrated electricity price forecasting method based on failure scenario detection. This second flowchart divides steps S1 to S6 into three stages:
[0042] During the offline training phase, steps S1 to S3 are executed to complete historical data collection and preprocessing, failure scenario labeling and classifier training, general model training and special model training.
[0043] In the online prediction phase, steps S4 to S5 are executed to complete real-time data input, feature extraction, failure scenario detection, general model prediction, special model prediction, weight calculation, weighted integration, final prediction output, confidence output, and early warning.
[0044] During the model update phase, step S6 is executed, forming a feedback loop through new data accumulation and performance evaluation. If the performance has decreased or if a new scenario has emerged, the model is retrained; otherwise, new data accumulation is returned.
[0045] Further, in step S1, the historical electricity market data includes at least one of the following: central dispatch load, new energy output, external power transmission, cross-sectional power flow, unit operating capacity, ramp-up capability, fuel price, and temperature. The bidding space is calculated as: central dispatch load - new energy output - external power transmission.
[0046] Furthermore, in step S2, the training scene classifier uses a random forest or XGBoost model, and the probability of belonging to each failure scene in the current time period is... Where i represents the scenario type, the extracted features include at least one of the following: cross-sectional load rate, load change rate, bidding spatial quantile, and prediction error, and the failure scenario includes at least one of the following four:
[0047] In a congested scenario, the load rate of the critical section exceeds the preset high load rate threshold and the duration exceeds the duration threshold.
[0048] In the ramp-up scenario, the net load change rate exceeds the preset percentage threshold of the system ramp-up capacity, and the occurrence occurs within the preset peak period.
[0049] In non-linear interval scenarios, the bidding space is either lower than the historical low quantile threshold or higher than the high quantile threshold.
[0050] In scenarios where uncertainty is amplified, the absolute value of the prediction error for new energy sources exceeds the high quantile threshold of the historical distribution.
[0051] In a specific embodiment, the preset high load rate threshold can be 85%; the preset duration threshold can be 1 hour; the preset percentage threshold for system ramp-up capability can be 80%; the preset historical low quantile threshold can be the historical 5th quantile, the preset high quantile threshold can be the historical 95th quantile, and the preset historical distribution error high quantile threshold can be the historical 80th quantile.
[0052] Furthermore, in step S3, the general model and the special model use the same or different benchmark prediction models, including AutoGluon TabularPredictor.
[0053] In a specific embodiment, the aforementioned AutoGluon TabularPredictor is an open-source automated machine learning framework that can automatically complete the preprocessing of tabular data, training of various regression models (including XGBoost, LightGBM, CatBoost, neural networks, etc.), hyperparameter optimization, and model ensemble, ultimately outputting a regression model that can be directly used for prediction.
[0054] Furthermore, in step S4, the current real-time data includes at least one of the following: cross-sectional load rate, net load change rate, bidding spatial quantile, new energy prediction deviation, and time / season. Each scenario classifier includes a congestion scenario classifier, a ramp scenario classifier, a nonlinear interval classifier, and an uncertainty amplification classifier.
[0055] Figure 3 The structure diagram for defining failure scenarios and training detection models.
[0056] In practice, the aforementioned blocking scenario classifier, climbing scenario classifier, nonlinear interval classifier, and uncertainty amplification classifier are all binary classifiers trained by random forest classifiers. The current real-time data is input into the blocking scenario classifier, climbing scenario classifier, nonlinear interval classifier, and uncertainty amplification classifier to obtain the probability of the current time period belonging to each failure scenario. Specifically, the blocking scenario classifier, climbing scenario classifier, nonlinear interval classifier, and uncertainty amplification classifier can respectively obtain the blocking probability p_cong, climbing probability p_ramp, nonlinear probability p_nonlinear, and uncertainty probability p_uncertain.
[0057] Furthermore, in step S5, the formula for calculating the integrated weights is:
[0058] (1);
[0059] in, These correspond to four different failure scenarios. Corresponding to the general model, The confidence coefficient of the model. The prediction accuracy of each model on the validation set can be preset, or dynamically adjusted through online learning. With no probability of failure and Or normalization can be used. These represent the probabilities of four failure scenarios.
[0060] Furthermore, in step S5, the predicted values of each model include: the predicted value of the general model. Blocking-specific model predictions Predicted values from a hill-climbing model Nonlinear dedicated model prediction value and uncertainty-specific model predictions The weighted fusion calculation formula is as follows:
[0061] (2);
[0062] in, For the final electricity price forecast, These are the predicted values for each model.
[0063] In a specific embodiment, the predicted values of the above-mentioned general model Blockage-specific model prediction values Predicted values from a slope-climbing model Nonlinear dedicated model prediction values and uncertainty-specific model predictions The final electricity price forecast is obtained by weighting and merging the data with the integration weights. For details, please refer to [link / reference]. Figure 4 , Figure 4 The flowchart for adaptive ensemble prediction.
[0064] In another embodiment, in step S5, the integrated weights can also be learned by a meta-learner based on historical data, the meta-learner including linear regression or a neural network.
[0065] Furthermore, in step S6, retraining adopts a sliding window approach, using only the most recently updated data, or adopts an incremental learning approach to update the model parameters and retrain the corresponding scene classifier.
[0066] In a specific implementation, taking a week as an example, every Monday morning, the performance of each model is re-evaluated using the new data from the past week. If the mean absolute error (MAE) of a certain model in the corresponding scenario increases by more than 10% compared to the previous week, the model is retrained using the latest data. If a new failure mode is detected, such as the classifier discovering a new cluster, a new scenario can be defined manually.
[0067] Experimental verification example: Using two years of historical data from a provincial electricity market, with a time resolution of 1 hour, the confidence coefficients of the preset model are as follows: [The text then abruptly shifts to a different topic:] ...general model confidence coefficients... Confidence coefficient of the dedicated blocking model Confidence coefficient of hill climbing model Confidence coefficient of nonlinear dedicated model Confidence coefficient of uncertainty-specific model The method of the present invention was verified by using a single XGBoost model as the benchmark and MAE as the evaluation index. The experimental results show that the MAE of the present invention is reduced by 18% compared with the benchmark model on the overall test set, 32% in the congestion scenario, and 25% in the ramp scenario. The system effectively identifies high-risk periods and outputs early warning information.
[0068] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.
[0069] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.
Claims
1. An adaptive integrated electricity price forecasting method based on failure scenario detection, characterized in that, Includes the following steps: Step S1: Historical data collection and preprocessing, collecting historical electricity market data, calculating the bidding space, cleaning and normalizing the data, and dividing it into training set, validation set and test set; Step S2: Failure scenario definition and detection model training. Define at least one failure scenario, construct sample labels based on historical data, extract features, train a scenario classifier and use it to detect the probability of belonging to each failure scenario in real time. Step S3: Training the general model and the special model. Use all historical data to train the baseline prediction model, and for each failure scenario, select a subset of historical data for that scenario to train a special prediction model. Step S4: Real-time failure scenario detection. During online prediction, the current real-time data is obtained and input into each scenario classifier to obtain the probability of belonging to each failure scenario in the current time period. Step S5: Adaptive ensemble prediction. The ensemble weights of each model are dynamically calculated based on the scenario probability and the preset model confidence coefficient. The current features are input into each model to obtain the predicted values of each model. The predicted values are then weighted and fused to obtain the final electricity price prediction. The prediction confidence is calculated based on the scenario probability and the consistency of the model prediction. If the confidence is lower than the confidence threshold or the probability of any failure scenario is higher than the risk threshold, an early warning is generated. Step S6: Online model update. New data generated during the online prediction phase is collected periodically. The performance of the general model and each special model is evaluated. When the performance of any model drops below a preset threshold or a new failure scenario is detected, retraining is triggered, and the original model is replaced with the updated model.
2. The method according to claim 1, characterized in that, In step S1, the historical electricity market data includes at least one of the following: central dispatch load, new energy output, external power transmission, cross-sectional power flow, unit operating capacity, ramp-up capability, fuel price, and temperature. The bidding space is calculated as: central dispatch load - new energy output - external power transmission.
3. The method according to claim 1, characterized in that, In step S2, the training scene classifier uses a random forest or XGBoost model, and the probability of belonging to each failure scene in the current time period is... Where i represents the scenario type, the extracted features include at least one of the following: cross-sectional load rate, load change rate, bidding spatial quantile, and prediction error, and the failure scenario includes at least one of the following four: In a congested scenario, the load rate of the critical section exceeds the preset high load rate threshold and the duration exceeds the duration threshold. In the ramp-up scenario, the net load change rate exceeds the preset percentage threshold of the system ramp-up capacity, and the occurrence occurs within the preset peak period. In non-linear interval scenarios, the bidding space is either lower than the historical low quantile threshold or higher than the high quantile threshold. In scenarios where uncertainty is amplified, the absolute value of the prediction error for new energy sources exceeds the high quantile threshold of the historical distribution.
4. The adaptive integrated electricity price forecasting method based on failure scenario detection according to claim 1, characterized in that, In step S3, the general model and the special model use the same or different benchmark prediction models, including AutoGluon TabularPredictor.
5. The adaptive integrated electricity price forecasting method based on failure scenario detection according to claim 1, characterized in that, In step S4, the current real-time data includes at least one of the following: cross-sectional load rate, net load change rate, bidding spatial quantile, new energy prediction deviation, and time / season. Each scenario classifier includes a congestion scenario classifier, a ramp scenario classifier, a nonlinear interval classifier, and an uncertainty amplification classifier.
6. The method according to claim 1, characterized in that, In step S5, the formula for calculating the integration weight is: ； in, These correspond to four different failure scenarios. General model The confidence coefficient of the model. The prediction accuracy of each model on the validation set can be preset, or dynamically adjusted through online learning. With no probability of failure and Or normalization can be used. These represent the probabilities of four failure scenarios.
7. The method according to claim 1, characterized in that, In step S5, the predicted values of each model include: the predicted value of the general model. Blocking-specific model predictions Predicted values from a hill-climbing model Nonlinear dedicated model prediction value and uncertainty-specific model predictions The weighted fusion calculation formula is as follows: ； in, For the final electricity price forecast, These are the predicted values for each model.
8. The method according to claim 1, characterized in that, In step S5, the integrated weights can also be learned by a meta-learner based on historical data, and the meta-learner includes linear regression or neural networks.
9. The method according to claim 1, characterized in that, In step S6, retraining is performed using a sliding window approach, using only the most recently updated data, or using incremental learning to update the model parameters and retrain the corresponding scene classifier.