A multi-feature fusion time series prediction method, system, device and medium for power load peak

Abstract

This invention discloses a multi-feature fusion time series prediction method, system, device, and medium for power load peak prediction, belonging to the field of power system load prediction technology. It includes: acquiring and preprocessing raw power load time series data; extracting and constructing a load peak time series for feature construction to obtain a prediction feature set; constructing supervised learning samples to transform the time series prediction problem into a regression problem; and training the supervised learning samples using a machine learning model to obtain a load peak prediction model for predicting and evaluating future load peaks. The beneficial effects of this invention are as follows: Through a feature engineering method of multi-feature fusion, this invention can effectively capture the historical dependence, short-term volatility, and long-term periodicity in power load time series, thereby significantly improving the accuracy of load peak prediction. It has a predictive advantage, especially for power systems with large load fluctuations, in environments with large-scale integration of new energy sources.

Description

Technical Field

[0001] This invention relates to the field of power system load forecasting technology, specifically to a multi-feature fusion time series forecasting method, system, equipment, and medium for power load peak. Background Technology

[0002] With the continuous increase in the proportion of new energy installed capacity and the increasing complexity of electricity load structure, the reliance of power system operation on the accuracy of load forecasting is constantly increasing. Especially in key operations such as grid dispatching, capacity planning, and system security assessment, the results of load peak forecasting directly affect the configuration of grid equipment, the arrangement of operation modes, and the effectiveness of peak-shaving resource deployment. Accurately grasping the variation patterns of load peaks is of great significance for ensuring the safe and stable operation of the power grid.

[0003] Currently, power load forecasting methods mainly include statistical model-based forecasting methods and machine learning-based forecasting methods. Traditional statistical methods, such as linear regression and time series analysis models, have limited capabilities in handling the periodicity, nonlinearity, and abrupt changes in load data, making them difficult to adapt to the increased load volatility under conditions of high renewable energy penetration. With improved computing power, machine learning methods such as support vector machines, neural networks, and ensemble learning models are gradually being applied to the field of load forecasting, improving forecast accuracy to some extent.

[0004] However, existing load forecasting methods still generally suffer from the following problems: On the one hand, most methods focus on directly modeling the original load time series, failing to specifically model and optimize the load peaks that are of great concern to the power grid operation, resulting in insufficient ability to characterize extreme changes, peak loads, and abnormal load scenarios; on the other hand, existing methods mostly remain at the level of simple historical lag or a small number of statistical features at the feature construction level, making it difficult to systematically integrate historical load dependencies, short-term fluctuation characteristics, and multi-timescale periodic information, thereby limiting the model's ability to learn complex load evolution mechanisms.

[0005] Furthermore, existing methods generally lack a systematic design for constructing time series supervised learning samples. When faced with problems such as strong non-stationarity of load data, complex periodicity, and significant noise interference, the stability and engineering adaptability of the models are still insufficient, making it difficult to meet the comprehensive requirements of actual power grid operation for the accuracy, reliability, and interpretability of load peak prediction. Summary of the Invention

[0006] In view of the above-mentioned problems, the present invention is proposed.

[0007] Therefore, the technical problem solved by this invention is: how to overcome the problems of low prediction accuracy, poor model stability and weak engineering adaptability caused by insufficient ability to characterize load peaks, limited feature construction dimensions and imperfect time series supervised learning sample construction mechanism in the prior art.

[0008] To address the aforementioned technical problems, this invention provides the following technical solution: a multi-feature fusion time series prediction method for peak power load, comprising, Acquire and preprocess raw time-series power load data, and extract and construct peak load time series; Feature construction is performed based on the peak load time series to obtain a feature set for prediction; The feature set is constructed as a supervised learning sample, transforming the time series prediction problem into a regression problem; A load peak prediction model is obtained by training supervised learning samples using a machine learning model. The trained load peak prediction model is used to predict and assess future load peaks.

[0009] As a preferred embodiment of the multi-feature fusion time series prediction method for peak power load described in this invention, the step of acquiring and preprocessing the original time series data of power load, and extracting and constructing the peak load time series includes: The raw time-series data containing timestamps and corresponding load values ​​is obtained and preprocessed to obtain continuous and ordered load data; The preprocessed load data is grouped according to a preset time scale, and the maximum value extraction operation is performed. Based on the extracted maximum load value, a load peak time series is constructed.

[0010] As a preferred embodiment of the multi-feature fusion time series prediction method for peak power load described in this invention, wherein: the feature construction based on the peak load time series to obtain the prediction feature set includes, Based on the peak load time series, a time series feature reflecting historical dependencies is constructed; Based on the peak load time series, statistical features reflecting short-term fluctuation characteristics are constructed; Based on calendar time information, time-dimensional features reflecting cyclical changes are constructed.

[0011] This invention addresses the shortcomings of traditional methods, which rely on singular feature construction and cannot systematically characterize the complex evolution patterns of load peaks, by constructing a multi-dimensional feature set that integrates historical dependence, short-term fluctuations, and cyclical changes. This enables the prediction model to simultaneously learn long-term load trends, short-term abrupt changes, and multi-scale cyclical characteristics, improving the prediction accuracy of load peaks, especially extreme points and abnormal peaks, in scenarios with strong fluctuations caused by high penetration of new energy sources.

[0012] As a preferred embodiment of the multi-feature fusion time series prediction method for peak power load described in this invention, the step of constructing the feature set as supervised learning samples and transforming the time series prediction problem into a regression problem includes: Each value in the peak load time series is defined as the output target of a supervised learning sample; The set of features corresponding to the historical moment for each output target is defined as the input features; By combining input features and output targets, a supervised learning sample set is formed, thus completing the transformation from time series to regression problems.

[0013] This invention systematically transforms the time series prediction problem into a supervised learning regression problem and constructs explicit input-output sample pairs, thus solving the problems of model instability and poor engineering adaptability caused by imperfect sample construction mechanisms in traditional time series modeling. This invention enables full utilization of robust and efficient machine learning frameworks such as ensemble learning for modeling, enhancing the stability of the prediction model when facing complex noise and non-stationary data.

[0014] As a preferred embodiment of the multi-feature fusion time series prediction method for peak power load described in this invention, the step of training a supervised learning sample using a machine learning model to obtain a peak load prediction model includes: The input features in the supervised learning sample set are scaled to meet the numerical requirements of model training. An ensemble learning regression model is used to iteratively train the transformed samples by minimizing a preset target loss function; When the ensemble learning regression model training reaches convergence, the final load peak prediction model is obtained.

[0015] As a preferred embodiment of the multi-feature fusion time series prediction method for peak power load described in this invention, the step of predicting and evaluating future peak loads using a trained peak load prediction model includes: The input features corresponding to the period to be predicted are scaled and then input into the load peak prediction model to obtain the prediction results. Transform the prediction results to restore them to their original dimensions; The restored prediction results are evaluated using preset evaluation indicators. Output the final peak load forecast and the corresponding evaluation results.

[0016] As a preferred embodiment of the multi-feature fusion time series prediction method for peak power load described in this invention, the step of evaluating the restored prediction results using preset evaluation indicators includes: A multi-indicator evaluation system is constructed to comprehensively evaluate the prediction results. The evaluation indicators include mean squared error (MSE), mean absolute error (MAE), root mean square error (RMSE), and coefficient of determination. And mean absolute percentage error (MAPE); The prediction results and corresponding evaluation indicators are stored in a database or file system, and prediction trend charts, error distribution charts and feature importance analysis charts are generated to help operators understand the model's prediction behavior and decision-making basis.

[0017] This invention provides a multi-feature fusion time series prediction system for peak power load.

[0018] To solve the above-mentioned technical problems, the present invention provides the following technical solution: a multi-feature fusion time series prediction system for peak power load, comprising: a data collection module, a feature construction module, a sample construction module, a model construction module, and an output evaluation module; The data collection module acquires and preprocesses the raw time-series data of power load, and extracts and constructs the peak load time series. The feature construction module constructs features based on the load peak time series to obtain a set of features for prediction. The sample construction module constructs a supervised learning sample from the feature set, transforming the time series prediction problem into a regression problem. The model building module uses a machine learning model to train supervised learning samples to obtain a load peak prediction model; The output evaluation module uses the trained load peak prediction model to predict and evaluate future load peaks.

[0019] The present invention provides a computer device, including a memory and a processor, wherein the memory stores a computer program, characterized in that the processor executes the computer program to implement the steps of the multi-feature fusion time series prediction method for power load peak.

[0020] The present invention provides a 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 steps of the multi-feature fusion time series prediction method for power load peak.

[0021] The beneficial effects of this invention are as follows: Through the feature engineering method of multi-feature fusion, this invention can effectively capture the historical dependence, short-term volatility and long-term periodicity in the time series of power load, thereby significantly improving the accuracy of load peak prediction. It has a predictive advantage, especially for power systems with large load fluctuations, in the context of large-scale integration of new energy sources.

[0022] This invention combines multi-dimensional features, including hysteresis characteristics, rolling window characteristics, time-dimensional characteristics, and periodic coding characteristics, to comprehensively characterize the evolution of peak power load. This fusion of multi-dimensional features effectively improves the adaptability of the prediction model to different load patterns and enhances the model's robustness and stability when facing complex and fluctuating data.

[0023] This invention can accurately predict peak loads, providing reliable decision support for power grid dispatching, capacity planning, and system safety operation. By providing early warnings of peak load issues, the power grid can conduct more rational load dispatching and resource allocation, ensuring the safe and efficient operation of the power system, especially providing crucial protection in the face of emergencies and extreme loads.

[0024] The load peak prediction method of this invention can help power grids optimize load dispatching and energy storage system operation strategies, reduce reliance on reserve capacity and peak-shaving resources, and thus effectively reduce power grid operating costs. By accurately predicting load fluctuations, power grids can more flexibly dispatch energy, reduce excess power and power waste, and improve system economy.

[0025] This invention is not only applicable to peak load forecasting, but can also be widely applied to power system capacity planning, load management, and dispatch optimization. Its unique multi-feature fusion method and ensemble learning-based model architecture enable it to effectively cope with various complex power grid environments, making it particularly suitable for load forecasting and dispatching in future smart grids, microgrids, and renewable energy power systems.

[0026] This invention has strong technical flexibility and can adjust the construction method of feature engineering and the training parameters of the model according to the needs and data characteristics of different power systems.

[0027] Therefore, this invention can be effectively applied and exhibits excellent performance in both traditional power grid environments and new smart grids and distributed energy systems. Attached Figure Description

[0028] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying 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 The above is a flowchart of a multi-feature fusion time series prediction method for peak power load provided in one embodiment of the present invention.

[0030] Figure 2 This figure illustrates the power load prediction results of a multi-feature fusion time series prediction method for peak power load provided in an embodiment of the present invention.

[0031] Figure 3 This figure shows the prediction results of the training and testing sets of a prediction model for a multi-feature fusion time series prediction method for peak power load provided in an embodiment of the present invention.

[0032] Figure 4 Error distribution diagram of power load prediction results of a multi-feature fusion time series prediction method for power load peak provided in an embodiment of the present invention.

[0033] Figure 5 This is a feature importance ranking diagram for a multi-feature fusion time series prediction method for peak power load provided in one embodiment of the present invention. Detailed Implementation

[0034] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0035] Example 1, referring to Figure 1 This is one embodiment of the present invention, which provides a multi-feature fusion time series prediction method for peak power load, including: To address the common problems of traditional power load forecasting methods, such as limited feature construction dimensions, lack of targeted characterization of load peaks, and insufficient model stability due to unsystematic time series sample construction mechanisms, this invention provides a systematic solution. Traditional methods often directly model the original load sequence, failing to effectively integrate historical dependencies, short-term fluctuations, and multi-scale periodic information, and lacking a reliable supervised learning transformation mechanism, resulting in limited accuracy and engineering adaptability for peak prediction.

[0036] The method provided in this embodiment significantly enhances feature representation capabilities by specifically constructing load peak time series and implementing multi-dimensional feature fusion; furthermore, through systematic supervised learning sample construction, the time series prediction problem is transformed into a stable regression problem, and finally, high-precision and highly robust load peak prediction is achieved by using an ensemble learning model.

[0037] S1. Acquire and preprocess the raw time-series data of power load, and extract and construct the peak load time series; S2 constructs features based on the peak load time series to obtain a set of features for prediction; S3. Construct the feature set as supervised learning samples and transform the time series prediction problem into a regression problem; S4 uses a machine learning model to train supervised learning samples to obtain a load peak prediction model; S5. Use the trained load peak prediction model to predict and evaluate future load peaks.

[0038] Example 2, an embodiment of the present invention, provides a multi-feature fusion time series prediction method for peak power load based on the previous embodiment, including: S1. Acquiring and preprocessing the raw time-series data of power load, and extracting and constructing the peak load time series includes the following steps: S11. Obtain the original time-series data containing timestamps and corresponding load values, perform preprocessing, and obtain continuous and ordered load data.

[0039] Obtain raw power load time series data.

[0040] The raw data should include at least timestamp information and the corresponding load value at that time.

[0041] Preferably, the raw data comes from the power monitoring system, dispatching system, or historical operation database, and is stored in the form of spreadsheets or databases.

[0042] S12. Group the preprocessed load data according to the preset time scale and perform the maximum value extraction action.

[0043] S13. Based on the extracted maximum load value, construct the load peak time series.

[0044] Through the process of constructing daily load peaks, the original high-frequency load data is transformed into a sequence of key indicators that can directly reflect the risks and capacity demands of power grid operation. This enables the model to learn more effectively the evolution of load peaks during power system operation, thereby improving the relevance and engineering applicability of daily load peak prediction.

[0045] In the embodiments of this application, the preprocessing of S11 is to convert the time field into a unified time format, sort the data according to the chronological order, and process outliers and missing data; in the case of missing data, forward padding or interpolation is used to complete the data to ensure the continuity of the time series.

[0046] In one optional implementation, the preprocessing in S11 involves introducing a sliding window smoothing technique based on sorting and outlier handling. Specifically, a sliding window of width M (e.g., M=5) is applied to the load sequence, and the median of the data within the window is calculated to replace the original value of the center point. This suppresses short-term random noise and transient impulse interference in the data, thereby improving the smoothness and quality of the data while ensuring that the trend remains unchanged.

[0047] In another optional implementation, the preprocessing of S11 is: outlier detection and processing based on statistical distribution. First, the mean and standard deviation of the load sequence are calculated, and data points that exceed the mean ± K times the standard deviation (e.g., K=3) are identified as outliers. For missing values ​​and identified outliers, linear interpolation of adjacent valid data points before and after that time is used to complete the data, so as to ensure the temporal continuity and statistical rationality of the data.

[0048] In the implementation of this application, the maximum value extraction action of S12 is to group the preprocessed load data according to the time scale of natural days, and extract the maximum value of the load values ​​at all times in each natural day group as the load peak value of that day.

[0049] In one optional implementation, the maximum value extraction action of S12 is performed by using a rolling time window for peak extraction. A rolling window of length L hours (e.g., L=24) is defined and slides over the load sequence. Each time, the maximum value among all load values ​​within the rolling window is extracted as the load peak value represented at the end of the window.

[0050] In another optional implementation, the maximum value extraction action of S12 is performed by using a rolling time window for peak extraction. A rolling window of length L hours (e.g., L=24) is defined and slides over the load sequence. Each time, the maximum value among all load values ​​within the rolling window is extracted as the load peak value represented at the end of the window.

[0051] This invention addresses the problem of insufficient peak prediction targeting caused by the dispersed objectives of traditional methods by constructing a dedicated load peak time series. This allows the model to learn the evolution of load peaks more accurately, improving the consistency and applicability of the prediction results with actual engineering needs such as power grid dispatching and capacity planning.

[0052] Meanwhile, by performing systematic preprocessing on the original data, including unifying the time format, sorting the time sequence, and handling outliers and missing values, continuous and ordered high-quality data was obtained, which fundamentally eliminated the unstable factors in model training caused by data disorder, noise, and missing values.

[0053] Furthermore, the diverse preprocessing and peak extraction options proposed in this invention, such as smoothing based on sliding windows and peak calculation based on Top-N weighted averages, demonstrate the adaptability of the core idea of ​​this invention to different data characteristics and business scenarios. This verifies that the technical solution has good robustness and flexibility while maintaining the consistency of the core logic, and can be applied to various complex practical engineering environments.

[0054] S2 constructs a feature set for prediction based on the peak load time series, including the following steps: S21. Based on the peak load time series, construct time series features that reflect historical dependencies.

[0055] To characterize the dependence of peak load on historical load conditions, a multi-order lag feature is constructed for the peak load time series, specifically including: lag_1, lag_2, lag_3, lag_7, lag_14, lag_21 and lag_28, Wherein, ·lag_1 represents the peak daily load value 1 day before the current forecast time; lag_2 represents the peak daily load value 2 days before the current forecast time; lag_3 represents the peak daily load value 3 days before the current forecast time; lag_7 represents the peak daily load value 7 days before the current forecast time; lag_14 represents the peak daily load value 14 days before the current forecast time; lag_21 represents the peak daily load value 21 days before the current forecast time; and lag_28 represents the peak daily load value 28 days before the current forecast time.

[0056] By constructing the aforementioned lag characteristics, the historical dependence of daily peak power load can be characterized from multiple time scales, enabling the prediction model to learn the short-term memory effect and medium- to long-term impact characteristics existing in the load evolution process, thereby improving the accuracy and stability of peak load prediction.

[0057] S22. Based on the peak load time series, construct statistical features that reflect short-term fluctuation characteristics.

[0058] To reflect the changing trend and fluctuation characteristics of peak load over a short time period, a rolling window statistical feature is constructed on the peak load time series, including: rolling_mean_3, rolling_std_7, rolling_max_14 and rolling_min_21, Wherein, rolling_mean_3 represents the average value of the peak load within a 3-day time window based on the current forecast time; rolling_std_7 represents the standard deviation of the peak load within a 7-day time window based on the current forecast time; rolling_max_14 represents the maximum value of the peak load within a 14-day time window based on the current forecast time; and rolling_min_21 represents the minimum value of the peak load within a 21-day time window based on the current forecast time.

[0059] By constructing the aforementioned rolling statistical features, we can characterize the statistical variation patterns and fluctuations of daily peak loads in the short to medium term across multiple time scales, providing more comprehensive and stable input information for the prediction model.

[0060] S23. Based on calendar time information, construct time-dimensional features that reflect the cyclical change pattern.

[0061] To capture the periodicity of load peaks over calendar time, time-dimensional features are constructed, including: The weekday (day of week), month (month), quarter (quarter), day-of-month (day of year), and week-of-year (week of year) features are used to describe the periodic variation characteristics of peak load on weekly, monthly, quarterly, and annual scales. Subsequently, to address the discontinuity of discrete calendar features at periodic boundaries, trigonometric functions are used to periodically encode the weekday and month features to construct continuous representations of periodic information.

[0062] It should be noted that the use of trigonometric functions to periodically encode the weekday and month features addresses the discontinuity of time-dimensional features at the periodic boundaries. By periodically encoding some time-dimensional features using trigonometric functions, the following features are constructed: sin_day_of_week, cos_day_of_week, sin_month, and cos_month. This achieves continuous expression of time-cycle information and improves the model's ability to learn cyclical change patterns.

[0063] The weekday encoding, expressed as: Month encoding, the expression is: Where d represents the week number of the current sample (ranging from 1 to 7), and m represents the month number of the current sample (ranging from 1 to 12).

[0064] By using the aforementioned periodic encoding method, time features that originally had abrupt changes in numerical space (such as the boundary relationship between Sunday and Monday, or between December and January) are mapped into a continuous and smooth feature space representation. This enables the model to correctly learn the continuous change characteristics of time cycles at the boundaries, thereby avoiding misjudgment problems caused by abrupt changes in boundary values ​​and significantly improving the model's ability to model and predict the periodic change patterns of power load.

[0065] To enhance the model's ability to perceive the trend of load peak changes, differential features are constructed on the load peak time series, including diff_1 and diff_7, which represent the load peak changes between two adjacent days and between one week apart, respectively, to characterize the short-term trend and periodic variation of load peaks.

[0066] By collaboratively constructing the above multi-dimensional features, we can achieve a multi-faceted characterization of the peak power load evolution mechanism, providing a high-quality input feature foundation for subsequent prediction models.

[0067] In the implementation of this application, the feature construction of S2 is as follows: based on the constructed load peak time series, a feature engineering system with multi-feature fusion is constructed around the historical evolution characteristics, short-term fluctuation characteristics and multi-time scale periodic change patterns of power load, forming a feature set for prediction modeling.

[0068] In one optional implementation, feature construction for S2 involves building a simplified feature set across the three dimensions to reduce feature dimensionality and improve model training efficiency. Specifically, this includes constructing only key lag features (e.g., lag1lag1, lag7lag7, lag28lag28), constructing rolling statistical features (mean and standard deviation) for a single window (e.g., a 7-day window), and using only basic weekday and month features (without periodic encoding). This simplification reduces computational complexity while maintaining the core feature dimensions, making it suitable for scenarios with high real-time requirements.

[0069] In another alternative implementation, feature construction for S2 involves feeding the load peak time series into a one-dimensional convolutional neural network (1D-CNN) for automatic feature extraction. This CNN network, composed of multiple convolutional and pooling layers, automatically learns from the original peak sequence and outputs a set of high-level abstract feature vectors. These abstract feature vectors are functionally equivalent to a manually constructed set of fused features, serving as input features for subsequent supervised learning samples.

[0070] This invention addresses the problem of traditional methods having a single feature set that fails to fully characterize the evolution of peak load by constructing a feature set that integrates three dimensions: historical dependence, short-term fluctuations, and cyclical changes. This allows the model to simultaneously learn long-term load dependence, short-term fluctuations, and multi-scale cyclical patterns, improving the prediction accuracy of peak loads, especially extreme and abnormal peak values.

[0071] By systematically constructing a feature set that integrates three dimensions—historical dependence, short-term fluctuations, and cyclical changes—this method solves the problem that traditional methods have a single feature structure and are difficult to fully characterize the evolution of load peaks.

[0072] S3. Constructing the feature set into supervised learning samples and transforming the time series prediction problem into a regression problem includes the following steps: S31. Define each value in the peak load time series as the output target of a supervised learning sample.

[0073] In step S3, based on the multi-feature fusion feature set constructed in step S2, the peak load time series is reconstructed, transforming the time series prediction problem into a supervised learning problem.

[0074] Specifically, the peak load at each forecast time is used as the forecast target value.

[0075] S32. Define the set of features corresponding to the historical moments of each output target as input features.

[0076] The input feature vector is composed of the lag features, rolling statistical features, time dimension features, periodic coding features, differential features, and auxiliary features corresponding to each time node before the prediction time.

[0077] S33. By combining input features and output targets, a supervised learning sample set is formed, completing the transformation from time series to regression problems.

[0078] Construct supervised learning sample pairs, represented in the following form: in, This represents an input feature vector composed of multidimensional features. This represents the target load peak prediction at time t.

[0079] Through the above sample construction process, the original load peak time series {y1, y2, y3, ..., yt} is transformed into a sample set suitable for regression learning, enabling the prediction model to model and learn the load peak change pattern under a supervised learning framework, thereby significantly improving the model's fitting ability and prediction accuracy for complex time series characteristics.

[0080] S4 uses a machine learning model to train supervised learning samples to obtain a load peak prediction model, including the following steps: In step S4, the load peak prediction model is trained based on the supervised learning sample set constructed in step S3.

[0081] S41. Scale the input features in the supervised learning sample set to meet the numerical requirements for model training.

[0082] The constructed supervised learning sample set is subjected to integrity checks, and sample records containing missing values ​​are removed to ensure the validity and continuity of the training data. The number of training samples is also ensured to be no less than the preset minimum sample size, so as to improve the stability and reliability of model training.

[0083] To eliminate the impact of differences in feature units on the model training process, the input feature vector is standardized using the following formula: in, This indicates that the j-th feature value is obtained from the i-th sample. and Let these represent the mean and standard deviation of the j-th feature in the training sample set, respectively. These are the standardized feature values.

[0084] S42. An ensemble learning regression model is used to iteratively train the transformed samples by minimizing a preset target loss function.

[0085] An ensemble learning regression model based on gradient boosting decision trees is constructed as the load peak prediction model. It iteratively corrects the prediction residuals by progressively constructing multiple regression decision trees. The overall objective function of the model is expressed as: in, This represents the overall objective loss of the model. This represents the actual peak load of the sample. These are the model's predicted values. This is a sample-level loss function used to measure prediction error. N represents the number of regression trees that make up the model, and N represents the number of training samples. Indicates the first A regression decision tree, This is a model complexity regularization term used to suppress model complexity and prevent overfitting. By minimizing the above loss function By gradually optimizing the structure and parameters of each regression tree, the model can maintain good generalization ability while ensuring prediction accuracy, thereby constructing a stable and reliable peak power load prediction model.

[0086] In a preferred embodiment, the prediction model is implemented using an XGBoost regression model, and its prediction output is expressed as follows: in, This represents the function space of the regression tree. During model training, the convergence speed and prediction accuracy are improved by jointly optimizing the first and second gradients of the objective function.

[0087] The XGBoost model is preferably configured with the following parameters: 150 decision trees, a maximum depth of 6, a learning rate of 0.1, a sample sampling ratio of 0.8, and a feature sampling ratio of 0.8, in order to achieve a balance between model complexity and generalization performance.

[0088] S43. When the ensemble learning regression model training reaches convergence, the final load peak prediction model is obtained.

[0089] During model training, the model parameters are adjusted and verified until the model training error converges and meets the preset accuracy requirements, thereby obtaining the final load peak prediction model.

[0090] In the implementation of this application, the training of supervised learning samples by the machine learning model in S4 is carried out by using an ensemble learning regression model based on gradient boosting decision trees (such as XGBoost).

[0091] The input feature vectors in the supervised learning sample set are standardized by using the mean and standard deviation of each feature dimension of the training set to transform the feature values ​​into a distribution with a mean of 0 and a standard deviation of 1, so as to eliminate the influence of the unit.

[0092] An XGBoost regression model is constructed, which is trained iteratively on standardized samples by minimizing an objective function consisting of a prediction error loss function (such as mean squared error) and a model complexity regularization term. In each iteration, a new decision tree is added to fit the current prediction residual, and the tree structure is updated by optimizing the first and second gradients of the objective function.

[0093] Set the model hyperparameters (e.g., 150 decision trees, maximum depth of 6, and learning rate of 0.1), and monitor the performance through the validation set during training until the model training error converges to a stable state, thereby obtaining the final load peak prediction model.

[0094] In one alternative implementation, the training of supervised learning samples using a machine learning model in S4 is performed using an ensemble learning model with automatically optimized hyperparameters.

[0095] During the model building phase, instead of using pre-set fixed hyperparameters, an automated hyperparameter optimization framework (such as cross-validation based on Bayesian optimization or grid search) is introduced. The system automatically searches and determines the optimal parameter combination within a pre-set hyperparameter space (such as the range of tree number, depth, and learning rate), and then uses this optimal combination to train the final model.

[0096] In another alternative implementation, S4 trains the supervised learning samples using a machine learning model by using a deep learning model (such as a Long Short-Term Memory Network (LSTM) or a Temporal Convolutional Network (TCN)).

[0097] A similar scaling transformation is performed on the input features. The feature vectors of the original independent samples are reconstructed into continuous sequence samples in chronological order to meet the sequential input requirements of deep learning models.

[0098] Construct a deep learning network suitable for regression tasks (e.g., a network containing LSTM layers and fully connected layers). The model is trained end-to-end on sequence data using the backpropagation algorithm to minimize the loss function (e.g., mean squared error) between predicted and true values.

[0099] Early stopping is used to prevent overfitting. Training is stopped when the model's performance on the validation set no longer improves, thus obtaining the final load peak prediction model.

[0100] This invention employs an ensemble learning model based on gradient boosting (such as XGBoost) and trains it by minimizing a target loss function that includes a regularization term. This enables the model to accurately learn the complex nonlinear mapping relationship between multi-dimensional fused features and load peaks in a high-performance, highly parallel, and efficient manner, while effectively controlling model complexity and avoiding overfitting. This not only improves the final accuracy and reliability of load peak prediction but also ensures the model's robust stability when facing fluctuating data in real-world applications.

[0101] S5. Predicting and evaluating future load peaks using the trained load peak prediction model includes the following steps: S51. Input the input features corresponding to the period to be predicted, after scaling transformation, into the load peak prediction model to obtain the prediction results.

[0102] The input feature vector corresponding to the time period to be predicted is standardized in the same way as in the training phase and then input into the prediction model to obtain the predicted peak load result. S52. Transform the prediction results to restore them to their original dimensions.

[0103] The prediction results are denormalized to obtain the final peak load prediction value, and the result is output as a prediction sequence.

[0104] S53. Use preset evaluation indicators to evaluate the restored prediction results.

[0105] To comprehensively evaluate the performance of the prediction model, a multi-index evaluation system is constructed to comprehensively evaluate the prediction results. The evaluation indices include mean squared error (MSE), mean absolute error (MAE), root mean square error (RMSE), and coefficient of determination. And the Mean Absolute Percentage Error (MAPE), which is calculated using the following formula: in, This represents the actual peak load. To predict peak load, is the mean of the actual peak load, and N is the sample size.

[0106] S54. Output the final peak load prediction value and the corresponding evaluation results.

[0107] The prediction results and corresponding evaluation indicators are stored in a database or file system, and prediction trend charts, error distribution charts and feature importance analysis charts are generated to help operators understand the model's prediction behavior and decision-making basis.

[0108] Example 3, referring to Figure 2 and Figure 5 This invention provides a multi-feature fusion time series prediction method for peak power load, as one embodiment of the present invention. To verify the beneficial effects of the present invention, scientific demonstration is carried out through experiments.

[0109] This embodiment verifies the effectiveness of the method of the present invention in predicting daily load peak by using actual power load data from a certain area of ​​the Guizhou power grid. The specific steps are as follows: Data Source and Preprocessing: This embodiment uses power load data from December 1, 2022 to February 28, 2025, sourced from the Guizhou power grid system operation data. The raw data includes hourly power load values. First, this data is aggregated by calendar day, and the maximum load value within 24 hours of each day is extracted as the peak load for that day.

[0110] After data processing, the daily load peak time series was obtained, with a total of 821 sample points, covering the period from December 1, 2022 to February 28, 2025.

[0111] Feature Engineering and Model Training: Through feature engineering, a multi-dimensional feature set was constructed, including lag features, rolling window features, temporal features, and periodic encoding features, generating a total of 43 features. These features include: lag features (e.g., lag_1, lag_2), rolling mean and standard deviation (e.g., rolling_mean_3, rolling_std_7), temporal features (e.g., day_of_week, month), and periodic encoding features (e.g., sin_day_of_week, cos_day_of_week). Based on these features, supervised learning samples were constructed and trained using an XGBoost regression model. The training set consisted of data from the first 734 days, and the test set consisted of data from the last 59 days.

[0112] The model parameters were set as follows: 150 trees, maximum depth of 6, learning rate of 0.1, sample ratio of 0.8, and feature ratio of 0.8. The model was trained using the training set and used to make predictions on the test set.

[0113] The prediction results show that the method of the present invention has good prediction accuracy on the test set, with a MAPE of 5.62% and an R² of 0.9399. Figure 2 The graph shows a comparison between the model's predictions and the actual peak load. The green curve represents the actual value, and the red square represents the XGBoost prediction value, clearly demonstrating the model's prediction accuracy across multiple time periods.

[0114] Figure 3 A detailed comparison of XGBoost prediction results is presented, with the figures showing predictions for both the training and test sets. The prediction results from the training set data are very close to the actual load peaks, indicating that the model has been able to fit the load change patterns well during training. The test set data also demonstrates the model's good predictive ability, illustrating the effectiveness of the method in practical applications.

[0115] Figure 4 The distribution of prediction errors is shown in the figure. This figure allows for a visual observation of the distribution of prediction errors. The vast majority of errors are concentrated in a small range, which verifies the high accuracy of the method of the present invention in peak load prediction.

[0116] Figure 5 This figure illustrates the relative importance ranking of each feature during the XGBoost model training process. It reveals the degree to which each feature contributes to the prediction results, particularly showing that lag features and rolling statistics have a significant impact on the model's predictions, further demonstrating the advantages of this invention in feature engineering construction.

[0117] As can be seen from the above embodiments, the method of the present invention can accurately predict the daily peak power load, especially when facing load data with strong periodicity and volatility, the model exhibits high robustness and stability. This demonstrates the potential of the present invention in practical applications such as power grid dispatching, capacity planning, and system safe operation, particularly its application prospects in large-scale integration of new energy sources and complex power grid environments.

[0118] Example 4 is an embodiment of the present invention, which provides a multi-feature fusion time series prediction system for peak power load, including a data collection module, a feature construction module, a sample construction module, a model construction module, and an output evaluation module; The data collection module acquires and preprocesses the raw time-series data of power load, and extracts and constructs the peak load time series. The feature construction module constructs features based on the load peak time series to obtain a set of features for prediction. The sample construction module constructs a supervised learning sample from the feature set, transforming the time series prediction problem into a regression problem. The model building module uses a machine learning model to train supervised learning samples to obtain a load peak prediction model; The output evaluation module uses the trained load peak prediction model to predict and evaluate future load peaks.

[0119] This embodiment also provides an electronic device applicable to a multi-feature fusion time series prediction method for power load peaks, comprising: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the multi-feature fusion time series prediction method for power load peaks as proposed in the above embodiment.

[0120] This embodiment also provides a storage medium on which a computer program is stored. When the program is executed by a processor, it implements a multi-feature fusion time series prediction method for power load peak as proposed in the above embodiment.

[0121] The storage medium proposed in this embodiment and the method for multi-feature fusion time series prediction for power load peak proposed in the above embodiments belong to the same inventive concept. Technical details not described in detail in this embodiment can be found in the above embodiments, and this embodiment has the same beneficial effects as the above embodiments.

[0122] Based on the above description of the implementation methods, those skilled in the art can clearly understand that the present invention can be implemented using software and necessary general-purpose hardware, and of course, it can also be implemented using hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as a computer floppy disk, read-only memory (ROM), random access memory (RAM), flash memory, hard disk, or optical disk, etc., including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods of the various embodiments of the present invention.

[0123] 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 preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A multi-feature fusion time series prediction method for peak power load, characterized in that: include, Acquire and preprocess raw time-series power load data, and extract and construct peak load time series; Feature construction is performed based on the peak load time series to obtain a feature set for prediction; The feature set is constructed as a supervised learning sample, transforming the time series prediction problem into a regression problem; A load peak prediction model is obtained by training supervised learning samples using a machine learning model. The trained load peak prediction model is used to predict and assess future load peaks.

2. The multi-feature fusion time series prediction method for peak power load as described in claim 1, characterized in that: The acquisition and preprocessing of raw time-series power load data, and the extraction and construction of peak load time series, include: The raw time-series data containing timestamps and corresponding load values ​​is obtained and preprocessed to obtain continuous and ordered load data; The preprocessed load data is grouped according to a preset time scale, and the maximum value extraction operation is performed. Based on the extracted maximum load value, a load peak time series is constructed.

3. The multi-feature fusion time series prediction method for peak power load as described in claim 2, characterized in that: The feature set for prediction obtained by constructing features based on the load peak time series includes, Based on the peak load time series, a time series feature reflecting historical dependencies is constructed; Based on the peak load time series, statistical features reflecting short-term fluctuation characteristics are constructed; Based on calendar time information, time-dimensional features reflecting cyclical changes are constructed.

4. The multi-feature fusion time series prediction method for peak power load as described in claim 3, characterized in that: The step of constructing a feature set as supervised learning samples and transforming the time-series prediction problem into a regression problem includes... Each value in the peak load time series is defined as the output target of a supervised learning sample; The set of features corresponding to the historical moment for each output target is defined as the input features; By combining input features and output targets, a supervised learning sample set is formed, thus completing the transformation from time series to regression problems.

5. The multi-feature fusion time series prediction method for peak power load as described in claim 4, characterized in that: The method of training a machine learning model on supervised learning samples to obtain a load peak prediction model includes: The input features in the supervised learning sample set are scaled to meet the numerical requirements of model training. An ensemble learning regression model is used to iteratively train the transformed samples by minimizing a preset target loss function; When the ensemble learning regression model training reaches convergence, the final load peak prediction model is obtained.

6. The multi-feature fusion time series prediction method for peak power load as described in claim 5, characterized in that: The method of predicting and evaluating future load peaks using a trained load peak prediction model includes: The input features corresponding to the period to be predicted are scaled and then input into the load peak prediction model to obtain the prediction results. Transform the prediction results to restore them to their original dimensions; The restored prediction results are evaluated using preset evaluation indicators. Output the final peak load forecast and the corresponding evaluation results.

7. The multi-feature fusion time series prediction method for peak power load as described in claim 6, characterized in that: The evaluation of the restored prediction results using preset evaluation indicators includes: A multi-indicator evaluation system is constructed to comprehensively evaluate the prediction results. The evaluation indicators include mean squared error (MSE), mean absolute error (MAE), root mean square error (RMSE), and coefficient of determination. And mean absolute percentage error (MAPE); The prediction results and corresponding evaluation indicators are stored in a database or file system, and prediction trend charts, error distribution charts and feature importance analysis charts are generated to help operators understand the model's prediction behavior and decision-making basis.

8. A multi-feature fusion time series prediction system for peak power load, employing the multi-feature fusion time series prediction method for peak power load as described in any one of claims 1 to 7, characterized in that, include: The system includes a data collection module, a feature construction module, a sample construction module, a model construction module, and an output evaluation module. The data collection module acquires and preprocesses the raw time-series data of power load, and extracts and constructs the peak load time series. The feature construction module constructs features based on the load peak time series to obtain a set of features for prediction. The sample construction module constructs a supervised learning sample from the feature set, transforming the time series prediction problem into a regression problem. The model building module uses a machine learning model to train supervised learning samples to obtain a load peak prediction model; The output evaluation module uses the trained load peak prediction model to predict and evaluate future load peaks.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the multi-feature fusion time series prediction method for power load peak as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the multi-feature fusion time series prediction method for power load peak as described in any one of claims 1 to 7.

Images