An integrated learning-based power load probability prediction method and device

Abstract

The application provides a power load probability prediction method and device based on ensemble learning. The method comprises the following steps: S1, preprocessing historical load data and meteorological data, extracting features and screening feature subsets; S2, training multiple gradient boosting decision tree learners based on different feature subsets to generate multiple ensemble prediction models; S3, performing robust statistical aggregation on the prediction results of the multiple ensemble prediction models, processing extreme values by using the truncated mean and the Winsorized mean, and obtaining aggregated prediction values; and S4, calculating a prediction interval and a probability distribution based on the aggregated prediction values, and outputting a probability prediction result of the power load. By training multiple gradient boosting decision tree learners on different feature subsets and combining statistical average and stacked ensemble meta-learners, probability prediction is realized, the prediction accuracy and robustness are improved, the prediction variance is reduced, and probability prediction is realized.

Description

Technical Field

[0001] This application relates to the field of load forecasting technology, and in particular to a method and apparatus for probabilistic forecasting of power load based on ensemble learning. Background Technology

[0002] Short-term power load forecasting serves as a crucial foundation for the safe and economical operation and dispatch planning of power systems, and is widely applied in areas such as power grid dispatching, market trading, and risk management. Among related technologies, a deterministic point forecasting system has been constructed through the collaborative operation of time series analysis, classical regression models, and single machine learning models. Specifically, this system covers the entire process from historical load data collection and meteorological factor fusion to model training and output, including key aspects such as feature engineering, model fitting, and error assessment.

[0003] However, existing load forecasting methods, which directly employ a single gradient boosting decision tree model for point forecasting, fail to effectively characterize the highly nonlinear and uncertain characteristics of load influenced by meteorological factors, social activities, and random disturbances. Because single models are susceptible to the distribution of training samples, hyperparameter selection, and feature engineering, they exhibit high prediction variance and struggle to fully capture the complex and diverse dependencies between load and external factors. This can lead to inaccurate risk assessment, difficulties in optimizing reserve capacity, and failure of probabilistic power flow analysis, thereby impacting the economic efficiency and reliability of power system operation. Summary of the Invention

[0004] This application aims to at least partially address one of the technical problems in the related art.

[0005] Therefore, this application proposes a method, apparatus, electronic device, and storage medium for predicting the probability of power load based on ensemble learning.

[0006] One embodiment of this application proposes a power load probability prediction method based on ensemble learning, including:

[0007] S1 preprocesses historical load data and meteorological data, extracts time period features, load lag features and meteorological features, and selects feature subsets based on feature importance; S2, train multiple gradient boosting decision tree learners based on different feature subsets to generate multiple ensemble prediction models; S3, robustly aggregate the prediction results of the multiple integrated prediction models, and use truncated mean and Windsorized mean to process extreme values ​​to obtain aggregated prediction values; S4. Calculate the prediction interval and probability distribution based on the aggregated prediction values, and output the probability prediction result of the power load.

[0008] Optionally, the preprocessing of historical load data and meteorological data to extract time period features, load lag features, and meteorological features, and the selection of feature subsets based on feature importance, includes: Missing values ​​in the original load sequence are filled using linear interpolation or nearest neighbor mean, and abnormal load records are removed based on interquartile range or physical threshold. Extract the periodic features of hours, days of the week, and holidays from the timestamps to construct load lag terms, sliding window mean, and standard deviation; It integrates meteorological data of temperature, humidity, and wind speed, and enhances the expression of meteorological characteristics by accumulating temperature or cooling / heating days; Feature selection methods based on mutual information ranking and recursive feature elimination are used to filter feature subsets.

[0009] Optionally, the feature selection method based on mutual information ranking and recursive feature elimination to filter feature subsets includes: Calculate the mutual information value between each feature and the target load variable, and rank the features in descending order based on the mutual information value, retaining a preset number of features with the highest ranking. Based on the retained features, a recursive feature elimination method is used to iteratively remove the features that contribute the least to the model until a preset feature quantity threshold is reached, thus obtaining the final feature subset.

[0010] Optionally, the step of training multiple gradient boosting decision tree learners based on different feature subsets to generate multiple ensemble prediction models includes: For each feature subset, bootstrap sampling is performed on the training data to generate multiple bootstrap samples. Multiple gradient boosting decision tree learners are trained on each bootstrap sample. Each learner sequentially constructs a set of regression trees. Multiple ensemble prediction models are generated based on the negative gradient of the loss function fitted by each regression tree with respect to the prediction of the current model.

[0011] Optionally, the robust statistical aggregation of the prediction results from the multiple integrated prediction models, using truncated means and Windsorized means to process extreme values, to obtain aggregated prediction values ​​includes: The truncated mean is calculated by applying a truncated mean to the set of prediction results from multiple ensemble prediction models. The calculation results after truncating the mean are processed using Windsorized mean, and the Windsorized mean is calculated according to the formula as the aggregated prediction value.

[0012] Optionally, the step of calculating the prediction interval and probability distribution based on the aggregated prediction values ​​and outputting the probability prediction result of the power load includes: The dataset is divided into multiple folds in the outer loop. In each iteration, one fold is used as the validation set and the remaining folds constitute the training set. The inner loop is used for hyperparameter tuning. The gradient boosting decision tree model based on bootstrapping provides quantile estimation for prediction intervals and calculates root mean square error, coefficient of determination, successive ranking probability score, prediction interval coverage probability, prediction interval normalized average width, and Winkler score, outputting the probabilistic prediction results for power load.

[0013] Another embodiment of this application proposes a power load probability prediction device based on ensemble learning, comprising: The feature extraction module is used to preprocess historical load data and meteorological data, extract time period features, load lag features and meteorological features, and filter feature subsets based on feature importance; The model generation module is used to train multiple gradient boosting decision tree learners based on different feature subsets to generate multiple ensemble prediction models. The aggregation module is used to perform robust statistical aggregation on the prediction results of the multiple integrated prediction models, and uses truncated mean and Windsor mean to process extreme values ​​to obtain aggregated prediction values. The prediction module is used to calculate the prediction interval and probability distribution based on the aggregated prediction values, and output the probability prediction results of the power load.

[0014] Another embodiment of this application provides an electronic device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, it implements the method described in the foregoing aspect.

[0015] Another embodiment of this application proposes a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method described in the foregoing aspect.

[0016] Another embodiment of this application proposes a chip including processing circuitry configured to perform the method described in one aspect above.

[0017] Another embodiment of this application proposes a computer program product that, when executed by a processor, implements the method described in the foregoing aspect.

[0018] The power load probability prediction method, device, electronic device, chip, and storage medium based on ensemble learning proposed in this application can achieve the following beneficial effects: By integrating multiple gradient boosting decision tree learners and training them on different feature subsets, and combining statistical averaging and stacked ensemble meta-learners, probabilistic prediction is achieved, improving prediction accuracy and robustness, reducing prediction variance, and realizing probabilistic prediction.

[0019] Additional aspects and advantages of this application will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of this application. Attached Figure Description

[0020] The above and / or additional aspects and advantages of this application will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein: Figure 1 A flowchart illustrating a power load probability prediction method based on ensemble learning provided in an embodiment of this application; Figure 2 A schematic diagram of the structure of an ensemble learning-based power load probability prediction device provided in an embodiment of this application; Figure 3 This is a schematic diagram of the structure of a chip proposed in an embodiment of this application. Detailed Implementation

[0021] The embodiments of this application are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain this application, and should not be construed as limiting this application.

[0022] The following description, with reference to the accompanying drawings, describes an embodiment of an ensemble learning-based method, apparatus, electronic device, chip, and storage medium for predicting the probability of power load.

[0023] Figure 1 This is a flowchart illustrating a power load probability prediction method based on ensemble learning, provided in an embodiment of this application.

[0024] As one implementation, the power load probability prediction method based on ensemble learning in this application embodiment can be configured in an ensemble learning-based power load probability prediction device. This ensemble learning-based power load probability prediction device can be applied to any electronic device so that the electronic device can perform the ensemble learning-based power load probability prediction function.

[0025] Among them, electronic devices can be any device with computing capabilities, such as mobile terminals, which can be hardware devices with various operating systems, touch screens and / or displays, such as mobile phones, tablets, personal digital assistants, wearable devices, etc.

[0026] As another implementation, the power load probability prediction method based on ensemble learning in this application embodiment can also be executed by a chip with processing capabilities. The chip includes an image signal processing chip (ISP), a central processing unit (CPU), an application-specific integrated circuit (ASIC), a digital signal processor (DSP), a field-programmable gate array (FPGA), a system on a chip (SOC), a reduced instruction set computer (RISC), etc., which will not be listed here.

[0027] It should be noted that all data collection operations related to users in this application are conducted with the user's authorization and in strict compliance with relevant laws and regulations such as privacy and security.

[0028] like Figure 1 As shown, the method may include the following steps: S1 preprocesses historical load data and meteorological data, extracts time period features, load lag features and meteorological features, and selects feature subsets based on feature importance; In this embodiment, historical load data and meteorological data are preprocessed to eliminate noise and anomalies in the raw data and extract feature information that can effectively characterize load variation patterns. This preprocessing includes the detection and repair of missing and outlier values ​​to ensure the integrity and reliability of the data sequence. Based on this, features reflecting periodic load fluctuations are extracted from the time dimension, such as time-series attributes related to daily cycles, weekly cycles, and holidays. Simultaneously, lag features capable of characterizing historical load dependencies are constructed, including historical load values ​​and their statistics at different time scales, to capture the dynamic evolution trend of the load.

[0029] Furthermore, external meteorological factors strongly correlated with the load, such as temperature, humidity, and wind speed, are incorporated, and their nonlinear hysteresis relationship with the load is considered, thereby enriching the feature space. For categorical variables, a coding method suitable for tree models is used to convert them into numerical form.

[0030] Subsequently, the features were screened based on the feature importance assessment method, and a subset of features that contribute more to the prediction task was selected from the original feature set to reduce feature redundancy and overfitting risk and improve the generalization ability of the model.

[0031] Through the above preprocessing and feature engineering, data quality can be effectively improved, the ability of features to express load change patterns can be enhanced, and the risk of model overfitting can be reduced. This provides high-quality, high-information input for the training of subsequent prediction models, thereby improving the accuracy and robustness of power load probability prediction.

[0032] S2, train multiple gradient boosting decision tree learners based on different feature subsets to generate multiple ensemble prediction models; In this embodiment, multiple gradient boosting decision tree learners are trained based on multiple different feature subsets determined after feature selection, to generate multiple ensemble prediction models. Specifically, gradient boosting decision tree, as an ensemble learning method based on regression decision trees, is based on the sequential construction of a set of regression trees, where each new tree fits the negative gradient of the loss function with respect to the current model's prediction, thereby gradually correcting the prediction error in difficult-to-handle regions of the feature space.

[0033] Since different feature subsets can capture different aspects of the dependency patterns between load behavior and external factors such as weather, training learners based on these subsets can generate multiple prediction models with complementary information, thereby effectively improving the diversity of ensemble predictions. As one implementation method, multiple feature subsets can be selected from the complete feature set using mutual information ranking and recursive feature elimination. These subsets are then used to train different types of gradient boosting decision tree learners, such as LightGBM, XGBoost, and CatBoost. Each learner is trained on its corresponding feature subset, ultimately forming a set of ensemble prediction models capable of characterizing load change patterns from different perspectives. This step utilizes the differences in feature subsets to construct diverse learners, providing a rich and complementary foundation of prediction information for subsequent prediction aggregation.

[0034] By training multiple gradient boosting decision tree learners based on different feature subsets, it is possible to effectively reduce the prediction variance of a single model caused by differences in training sample distribution or feature selection, enhance the model's ability to fit complex nonlinear relationships, and improve the robustness of the overall prediction framework to overfitting under specific conditions.

[0035] S3, robustly aggregate the prediction results of the multiple integrated prediction models, and use truncated mean and Windsorized mean to process extreme values ​​to obtain aggregated prediction values; In this embodiment, after obtaining the prediction results from multiple ensemble prediction models, it is necessary to effectively aggregate these predictions from different feature subsets or different learner structures to generate a comprehensive final prediction result with higher stability and accuracy. The core of this aggregation process lies in employing robust statistical methods to overcome the potential biases in single-model predictions and the impact of extreme values ​​caused by data perturbations or model overfitting.

[0036] Specifically, this method statistically processes the set of predicted values ​​output by multiple ensemble prediction models, using two robust statistics—truncated mean and Windsorized mean—to replace the simple arithmetic average. The truncated mean eliminates the interference of extreme outliers on the average result by removing a certain proportion of the maximum and minimum values ​​at the ends of the predicted value sequence. The Windsorized mean, however, does not directly remove extreme values ​​but replaces them with the maximum or minimum values ​​of the remaining portion of the sequence, thus suppressing the impact of extreme values ​​while preserving sample size. Both methods effectively reduce the fluctuation of the aggregated results caused by inaccurate predictions from individual models or data noise, improving the representativeness of the aggregated predicted values ​​to the overall load change trend. Through the calculation of these two robust statistics, the final aggregated predicted value is obtained, which comprehensively reflects the prediction information of each model and is robust to outliers.

[0037] By aggregating the outputs of multiple integrated prediction models using robust statistical methods such as truncated mean and Windsorized mean, the robustness and stability of the final prediction results can be significantly improved. This effectively suppresses prediction errors caused by prediction biases of individual models or extreme noise in the data, thereby enhancing the adaptability of the prediction method to complex and variable power load environments while maintaining prediction accuracy.

[0038] S4. Calculate the prediction interval and probability distribution based on the aggregated prediction values, and output the probability prediction result of the power load.

[0039] In this embodiment, after obtaining the aggregated predicted value, a probability distribution that can characterize the prediction uncertainty is further constructed based on the aggregated predicted value, and the corresponding prediction interval is calculated accordingly. Specifically, by performing statistical distribution modeling on the predicted values ​​output by multiple ensemble prediction models, for example, by using the quantile estimates of multiple gradient boosting decision tree models generated by the bootstrap sampling method, a cumulative distribution function of the prediction results is constructed, thereby achieving a probabilistic description of the possible range of future load values.

[0040] Based on the constructed probability distribution, a prediction interval at a given confidence level can be calculated. This interval consists of a lower bound and an upper bound, used to quantify the range of uncertainty in the prediction result. Furthermore, this probability distribution can also output a complete probability density function or cumulative distribution function, thus providing richer information than a single deterministic prediction.

[0041] This step extends the prediction from deterministic to probabilistic by converting aggregated forecasts into probability distributions and calculating forecast intervals. It can effectively quantify the uncertainty in load forecasting and provide a more comprehensive and reliable basis for decision-making in applications such as power system risk assessment, reserve capacity optimization, and probabilistic power flow analysis.

[0042] Optionally, the preprocessing of historical load data and meteorological data to extract time period features, load lag features, and meteorological features, and the selection of feature subsets based on feature importance, includes: Missing values ​​in the original load sequence are filled using linear interpolation or nearest neighbor mean, and abnormal load records are removed based on interquartile range or physical threshold. Extract the periodic features of hours, days of the week, and holidays from the timestamps to construct load lag terms, sliding window mean, and standard deviation; It integrates meteorological data of temperature, humidity, and wind speed, and enhances the expression of meteorological characteristics by accumulating temperature or cooling / heating days; Feature selection methods based on mutual information ranking and recursive feature elimination are used to filter feature subsets.

[0043] In this embodiment, for missing and outlier values ​​in the original load sequence, the system receives the original load data as input, fills in the missing points using linear interpolation or nearest neighbor mean, and identifies and removes abnormal load records caused by metering failures or sudden events based on interquartile range or physical threshold, and outputs a cleaned load sequence.

[0044] Subsequently, time periodic features were extracted from the cleaned load sequence and corresponding timestamp data, specifically including periodic features of hours, days of the week, and holidays; at the same time, load lag features were constructed, including load lag terms of the previous 24 hours and the previous 168 hours, as well as the mean and standard deviation of the sliding window of the past 1 hour and 24 hours. These features together constitute a set of features related to time and historical load.

[0045] Based on this, the system further integrates meteorological data such as temperature, humidity, and wind speed, and enhances the expression of meteorological characteristics by accumulating temperature or cooling / heating days, so as to capture the nonlinear hysteresis relationship between meteorological factors and load, and outputs a comprehensive feature set containing meteorological information.

[0046] Finally, the comprehensive feature set is filtered based on the feature selection method of mutual information ranking and recursive feature elimination.

[0047] Through the above specific implementation methods, this method can effectively handle noise and missing data in the original data, fully mine key information in time, load history and meteorological factors, and significantly reduce feature redundancy and overfitting risk by using a feature selection strategy that combines mutual information ranking and recursive feature elimination, thereby improving the accuracy and robustness of the subsequent integrated prediction model.

[0048] Optionally, the feature selection method based on mutual information ranking and recursive feature elimination to filter feature subsets includes: Calculate the mutual information value between each feature and the target load variable, and rank the features in descending order based on the mutual information value, retaining a preset number of features with the highest ranking. Based on the retained features, a recursive feature elimination method is used to iteratively remove the features that contribute the least to the model until a preset feature quantity threshold is reached, thus obtaining the final feature subset.

[0049] In this embodiment, the mutual information value between each feature and the target variable is first calculated, and the features are ranked in descending order based on this value, retaining a preset number of features with the highest ranking. Then, based on the retained features, a recursive feature elimination method is used to iteratively remove features that contribute the least to the model until a preset feature quantity threshold is reached, thus obtaining the final feature subset. This feature subset serves as the input for training multiple gradient boosting decision tree learners in subsequent step S2, ensuring that the model can make predictions based on the most informative features.

[0050] Optionally, the step of training multiple gradient boosting decision tree learners based on different feature subsets to generate multiple ensemble prediction models includes: For each feature subset, bootstrap sampling is performed on the training data to generate multiple bootstrap samples. Multiple gradient boosting decision tree learners are trained on each bootstrap sample. Each learner sequentially constructs a set of regression trees. Multiple ensemble prediction models are generated based on the negative gradient of the loss function fitted by each regression tree with respect to the prediction of the current model.

[0051] In this embodiment, for each selected feature subset, the Bagging method is used to perform bootstrap sampling on the training data corresponding to that feature subset, generating multiple bootstrap samples. Specifically, for each feature subset, samples of the same size as the original dataset are randomly drawn with replacement from the original training dataset, forming B bootstrap samples, where B is the number of bootstrap samples. The variance is reduced by averaging the prediction results of the model trained on multiple bootstrap (sampling) samples.

[0052] in, The number of samples for the self-help method. For the first The predicted values ​​of each model.

[0053] Subsequently, three different types of gradient boosting decision tree learners—LightGBM, XGBoost, and CatBoost—were trained on each bootstrap sample. The regression decision tree divides the feature space into non-overlapping regions. And assign a constant prediction value to each region. The goal of a standard regression tree is to find the optimal split. To minimize the total squared error:

[0054] in, This represents the decision tree before pruning. For the number of terminal nodes, This is a complexity control parameter.

[0055] Each learner is trained by sequentially constructing a set of regression trees, where each tree fits the negative gradient of the loss function with respect to the current model's prediction. Specifically, for the m-th regression tree, according to the formula... Iteratively update the model, where For learning rate, Let M be the m-th regression tree, and M be the total number of trees. This iterative process completes model training by progressively correcting errors in difficult-to-handle regions of the feature space. That is, each new tree focuses on the sample regions where the previous model's prediction error was large, thereby gradually improving the overall model's fitting accuracy.

[0056] By independently training the three learners mentioned above on each bootstrap sample, multiple ensemble prediction models are generated for each feature subset. The prediction results of these models together constitute a prediction set, denoted as . Each of them This represents an integrated forecast. These forecasts capture different aspects of load behavior, providing diverse inputs for subsequent robust statistical aggregation.

[0057] This specific implementation significantly improves the diversity and robustness of the ensemble prediction model by introducing a combination of Bagging bootstrap sampling and multiple gradient boosting decision tree learners. It effectively reduces the prediction variance of a single model caused by the distribution of training samples or the selection of hyperparameters, thus providing a more reliable foundation for subsequent robust aggregation.

[0058] Optionally, the robust statistical aggregation of the prediction results from the multiple integrated prediction models, using truncated means and Windsorized means to process extreme values, to obtain aggregated prediction values ​​includes: The truncated mean is calculated by applying a truncated mean to the set of prediction results from multiple ensemble prediction models. The calculation results after truncating the mean are processed using Windsorized mean, and the Windsorized mean is calculated according to the formula as the aggregated prediction value.

[0059] In this embodiment, the prediction results of multiple gradient boosting decision tree learners trained based on different feature subsets in step S2 are first obtained. These prediction results constitute a set of predicted values, denoted as . Each of them This represents the output values ​​of an ensemble prediction model at the same prediction time on a specific subset of features. The individual predictions in this set capture different aspects of load behavior but may contain extreme values ​​due to overfitting the model on that specific subset of features. To eliminate the influence of these extreme values, a truncated mean is first applied. Specifically, the set... The predicted values ​​are sorted by numerical value to obtain the order statistic. ,in This represents the total number of predicted values.

[0060] Set a truncation parameter This parameter determines the number of extreme values ​​to remove from both ends of the sorted sequence. The smallest value is then removed. The largest sum of values Only the middle part of the values ​​is retained. One ordinal statistic, namely to The arithmetic mean of these retained values ​​is calculated to obtain the truncated mean, and the formula is as follows: The truncated mean, as the aggregation result of the first stage, effectively eliminates the interference of extreme predicted values ​​on the overall mean.

[0061] Next, to further utilize the information carried by the removed extreme values ​​and avoid information loss due to simple removal, the Windsorized mean is used to correct the above processing results. Specifically, after obtaining the sorted ordinal statistics... to Afterwards, instead of directly removing extreme values, they are replaced: the smallest value is replaced. Value Replace all with the number Order statistics The value, while also the largest Value Replace all with the number Order statistics The value. After replacement, a new set of values ​​is obtained. ,in, Indicates the number after replacement A value. Finally, for this... The arithmetic mean of the replaced values ​​is calculated to obtain the Windsorized mean, which is calculated using the following formula: The Windsorized mean is the final aggregated prediction. It suppresses the influence of extreme values ​​while retaining the total sample size, making the aggregated results more robust and reliable.

[0062] This specific implementation method effectively suppresses the interference of individual extreme predictions in the ensemble prediction model on the final aggregation result through a two-step robust statistical processing of first truncation and then Windsorization. While preserving the diversity of predictions, it significantly improves the robustness and accuracy of the aggregated predictions.

[0063] Optionally, the step of calculating the prediction interval and probability distribution based on the aggregated prediction values ​​and outputting the probability prediction result of the power load includes: The dataset is divided into multiple folds in the outer loop. In each iteration, one fold is used as the validation set and the remaining folds constitute the training set. The inner loop is used for hyperparameter tuning. The gradient boosting decision tree model based on bootstrapping provides quantile estimation for prediction intervals and calculates root mean square error, coefficient of determination, successive ranking probability score, prediction interval coverage probability, prediction interval normalized average width, and Winkler score, outputting the probabilistic prediction results for power load.

[0064] In this embodiment, based on the aggregated prediction value output in step S3, a nested k-fold cross-validation framework is used to generate and evaluate the probability prediction results. Specifically, the complete dataset D is first divided into k disjoint folds (i.e., subsets), denoted as . Each fold The remaining k-1 folds form the training set, serving as the validation set for one outer iteration.

[0065] In each iteration of the outer loop, the multiple gradient boosting decision tree ensemble models trained in step S2 are retrained or fine-tuned using the training set. Hyperparameter tuning is performed in the inner loop, for example, by determining the optimal combination of parameters such as the learning rate, maximum tree depth, and minimum number of samples per leaf node through grid search or Bayesian optimization. The validation results of the inner loop are used to select the best hyperparameter configuration, while the outer loop is used to evaluate the model's generalization performance on unseen data. After all k iterations of the outer loop, the predicted values ​​on the k validation sets are obtained.

[0066] Subsequently, quantile estimates from a gradient boosting decision tree model based on bootstrap sampling are used to provide the prediction interval. Specifically, for each test sample, B samples are extracted from the training set using a bootstrap method, and B gradient boosting decision tree models are trained respectively to obtain B predicted values. The empirical distribution of these predicted values ​​can be used to estimate arbitrary quantiles; for example, the 2.5% and 97.5% quantiles can be used as the lower bound of the prediction interval at the 90% confidence level. and the Upper Realm Therefore, for each time point t, the output probability prediction result includes the predicted point value (such as the mean or median) and the prediction interval. , ].

[0067] To comprehensively evaluate the performance of probabilistic prediction, the following six metrics are calculated: Root mean square error is based on the formula Calculation, where These are actual observed values. is the predicted value for a point, and n is the total number of samples.

[0068] The coefficient of determination is based on the formula calculate, This is the mean of the observed values.

[0069] The probability score for consecutive rankings is based on the formula. Calculation, where To predict the cumulative distribution function, This is an indicator function. Lower consecutive ranking probability scores indicate better probability prediction accuracy.

[0070] Predicting interval coverage probability based on the formula The calculation is performed, where N is the total number of time points. A higher PICP value (closer to the nominal confidence level) indicates better calibration results.

[0071] The normalized average width of the prediction interval is based on the formula. Calculation, where This represents the range of the observed values. Lower... The value indicates a sharper prediction interval.

[0072] Winkler scores are based on the formula calculate, The significance level is indicated by a score of 0.1 (e.g., 0.1 corresponds to a 90% confidence interval). A lower Winkler score indicates a narrower prediction interval and better calibration.

[0073] These indicators characterize the accuracy, calibration, and sharpness of probabilistic prediction from different dimensions, and ultimately output the power load probabilistic prediction results, which include point prediction, prediction interval, and the above evaluation results.

[0074] This implementation combines nested cross-validation with bootstrap quantile estimation, effectively avoiding overfitting and providing a reliable prediction range. At the same time, it utilizes multi-indicator comprehensive evaluation to ensure the comprehensiveness and robustness of probability prediction.

[0075] To achieve the above embodiments, this application also proposes a power load probability prediction device based on ensemble learning.

[0076] Figure 2 This is a schematic diagram of the structure of an ensemble learning-based power load probability prediction device provided in an embodiment of this application.

[0077] like Figure 2 As shown, the device may include: The feature extraction module 210 is used to preprocess historical load data and meteorological data, extract time period features, load lag features and meteorological features, and filter feature subsets based on feature importance; The model generation module 220 is used to train multiple gradient boosting decision tree learners based on different feature subsets to generate multiple ensemble prediction models. The aggregation module 230 is used to perform robust statistical aggregation on the prediction results of the multiple integrated prediction models, and uses truncated mean and Windsor mean to process extreme values ​​to obtain aggregated prediction values. The prediction module 240 is used to calculate the prediction interval and probability distribution based on the aggregated prediction value and output the probability prediction result of the power load.

[0078] It should be noted that the foregoing explanation of the method embodiments also applies to the apparatus of this embodiment, and will not be repeated here.

[0079] To implement the above embodiments, this application also proposes a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the method described in the foregoing method embodiments.

[0080] To implement the above embodiments, this application also proposes a computer program product having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the method described in the foregoing method embodiments.

[0081] To implement the above embodiments, this application also proposes an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, it implements the method described in the foregoing method embodiments.

[0082] Figure 3 This is a schematic diagram of the structure of a chip according to an embodiment of this application. See also... Figure 3 The diagram shown is a schematic representation of the structure of chip 1100, but it is not limited to this.

[0083] Chip 1100 includes processing circuitry 1101, which is configured to perform any of the above methods.

[0084] In some embodiments, chip 1100 further includes one or more interface circuits 1102. Optionally, the interface circuit 1102 is connected to memory 1103, and the interface circuit 1102 can be used to receive signals from memory 1103 or other devices, and the interface circuit 1102 can be used to send signals to memory 1103 or other devices. For example, the interface circuit 1102 can read instructions stored in memory 1103 and send the instructions to processing circuit 1101.

[0085] In some embodiments, the interface circuit 1102 performs at least one of the communication steps such as sending and / or receiving in the above method, while the processing circuit 1101 performs other steps.

[0086] In some embodiments, the terms interface circuit, interface, transceiver pin, transceiver, etc., can be used interchangeably.

[0087] In some embodiments, chip 1100 further includes one or more memories 1103 for storing instructions. Optionally, all or part of the memories 1103 may be located outside of chip 1100.

[0088] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0089] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this application, "multiple" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0090] Any process or method description in the flowchart or otherwise herein can be understood as representing a module, segment, or portion of code comprising one or more executable instructions for implementing custom logic functions or processes, and the scope of the preferred embodiments of this application includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order depending on the functions involved, as should be understood by those skilled in the art to which embodiments of this application pertain.

[0091] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-included system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device. More specific examples (a non-exhaustive list) of computer-readable media include: an electrical connection having one or more wires (electronic device), a portable computer disk drive (magnetic device), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Alternatively, the computer-readable medium may be paper or other suitable media on which the program can be printed, since the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in a computer memory.

[0092] It should be understood that various parts of this application can be implemented using hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented using software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0093] Those skilled in the art will understand that all or part of the steps of the methods in the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, the program includes one or a combination of the steps of the method embodiments.

[0094] Furthermore, the functional units in the various embodiments of this application can be integrated into a processing module, or each unit can exist physically separately, or two or more units can be integrated into a module. The integrated module can be implemented in hardware or as a software functional module. If the integrated module is implemented as a software functional module and sold or used as an independent product, it can also be stored in a computer-readable storage medium.

[0095] The storage medium mentioned above can be a read-only memory, a disk, or an optical disk, etc. Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make changes, modifications, substitutions, and variations to the above embodiments within the scope of this application.

Claims

1. A power load probabilistic forecasting method based on ensemble learning, characterized in that, include: S1 preprocesses historical load data and meteorological data, extracts time period features, load lag features and meteorological features, and selects feature subsets based on feature importance; S2, train multiple gradient boosting decision tree learners based on different feature subsets to generate multiple ensemble prediction models; S3, robustly aggregate the prediction results of the multiple integrated prediction models, and use truncated mean and Windsorized mean to process extreme values ​​to obtain aggregated prediction values; S4. Calculate the prediction interval and probability distribution based on the aggregated prediction values, and output the probability prediction result of the power load.

2. The method according to claim 1, characterized in that, The process of preprocessing historical load data and meteorological data to extract time period features, load lag features, and meteorological features, and then selecting a feature subset based on feature importance, includes: Missing values ​​in the original load sequence are filled using linear interpolation or nearest neighbor mean, and abnormal load records are removed based on interquartile range or physical threshold. Extract the periodic features of hours, days of the week, and holidays from the timestamps to construct load lag terms, sliding window mean, and standard deviation; It integrates meteorological data of temperature, humidity, and wind speed, and enhances the expression of meteorological characteristics by accumulating temperature or cooling / heating days; Feature selection methods based on mutual information ranking and recursive feature elimination are used to filter feature subsets.

3. The method according to claim 2, characterized in that, The feature selection method based on mutual information ranking and recursive feature elimination filters feature subsets, including: Calculate the mutual information value between each feature and the target load variable, and rank the features in descending order based on the mutual information value, retaining a preset number of features with the highest ranking. Based on the retained features, a recursive feature elimination method is used to iteratively remove the features that contribute the least to the model until a preset feature quantity threshold is reached, thus obtaining the final feature subset.

4. The method according to claim 1, characterized in that, The process involves training multiple gradient boosting decision tree learners based on different feature subsets to generate multiple ensemble prediction models, including: For each feature subset, bootstrap sampling is performed on the training data to generate multiple bootstrap samples. Multiple gradient boosting decision tree learners are trained on each bootstrap sample. Each learner sequentially constructs a set of regression trees. Multiple ensemble prediction models are generated based on the negative gradient of the loss function fitted by each regression tree with respect to the prediction of the current model.

5. The method according to claim 4, characterized in that, The robust statistical aggregation of the prediction results from the multiple integrated prediction models, using truncated mean and Windsorized mean to process extreme values, yields aggregated prediction values, including: The truncated mean is calculated by applying a truncated mean to the set of prediction results from multiple ensemble prediction models. The calculation results after truncating the mean are processed using Windsorized mean, and the Windsorized mean is calculated according to the formula as the aggregated prediction value.

6. The method according to claim 1, characterized in that, The step of calculating the prediction interval and probability distribution based on the aggregated prediction values ​​and outputting the probability prediction result of the power load includes: The dataset is divided into multiple folds in the outer loop. In each iteration, one fold is used as the validation set and the remaining folds constitute the training set. The inner loop is used for hyperparameter tuning. The gradient boosting decision tree model based on bootstrapping provides quantile estimation for prediction intervals and calculates root mean square error, coefficient of determination, successive ranking probability score, prediction interval coverage probability, prediction interval normalized average width, and Winkler score, outputting the probabilistic prediction results for power load.

7. A power load probability prediction device based on ensemble learning, characterized in that, include: The feature extraction module is used to preprocess historical load data and meteorological data, extract time period features, load lag features and meteorological features, and filter feature subsets based on feature importance; The model generation module is used to train multiple gradient boosting decision tree learners based on different feature subsets to generate multiple ensemble prediction models. The aggregation module is used to perform robust statistical aggregation on the prediction results of the multiple integrated prediction models, and uses truncated mean and Windsor mean to process extreme values ​​to obtain aggregated prediction values. The prediction module is used to calculate the prediction interval and probability distribution based on the aggregated prediction values, and output the probability prediction results of the power load.

8. An electronic device, characterized in that, It includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, it implements the method as described in any one of claims 1-6.

9. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method as described in any one of claims 1-6.

10. A chip, characterized in that, The chip includes processing circuitry configured to perform the method as described in any one of claims 1-6.

Images