Short-term power load prediction method

Abstract

The invention provides a short-term power load prediction method. The method comprises the following steps: acquiring historical load data of a historical time period, historical meteorological data and historical time data corresponding to the historical load data, and future meteorological data and future time data of a to-be-predicted time period; according to the historical load data, the historical meteorological data and the historical time data, influence factor characteristics of the historical load data are obtained; according to the influence factor characteristics, determining influence factor data from historical meteorological data, historical time data, future meteorological data and future time data; performing modal decomposition on the historical load data to obtain a plurality of load components with different frequencies; according to a preset time sequence prediction model, the influence factor data and each load component, determining a prediction result corresponding to each load component; and combining and reconstructing the prediction results to obtain power load prediction data of the to-be-predicted time period. The method can improve the prediction precision of the short-term power load.

Description

Technical Field

[0001] The present application relates to the technical field of load forecasting, and in particular to a short-term power load forecasting method. Background Art

[0002] In the ongoing reform of the power system and the dynamic expansion of the power market, accurate short-term power load forecasting is essential to facilitate real-time transaction settlement within the power market. The inherent uncertainty surrounding power load often causes forecast differences, increases the operating costs of the power grid, and poses a major challenge to power system dispatching.

[0003] There are two main types of short-term power load forecasting in related technologies. One is the classical forecasting method, which uses support vector machines, artificial neural networks, extreme learning machines and their combined models to forecast short-term power loads. The other is the method that uses artificial intelligence technology such as tree ensemble algorithms and deep learning methods to forecast short-term power loads, which provides strong support for accurate decision-making in the power industry. However, most of these methods focus on point forecasting, which will significantly affect the accuracy of load forecasting in complex environments and often cannot achieve ideal results in long-term series forecasting. Summary of the invention

[0004] The embodiment of the present application provides a short-term power load forecasting method to improve the prediction accuracy and performance of short-term power load in complex environments.

[0005] In a first aspect, an embodiment of the present application provides a short-term power load forecasting method, comprising:

[0006] Obtain historical load data for a historical period, historical meteorological data and historical time data corresponding to the historical load data, and future meteorological data and future time data for a period to be predicted;

[0007] Performing feature analysis based on the historical load data, the historical meteorological data, and the historical time data to obtain the features of influencing factors of the historical load data;

[0008] Determining influencing factor data from the historical meteorological data, the historical time data, the future meteorological data and the future time data according to the influencing factor characteristics;

[0009] Performing modal decomposition on the historical load data to obtain multiple load components of different frequencies;

[0010] Determine the prediction result corresponding to each load component according to the preset time series prediction model, influencing factor data and each load component;

[0011] The various prediction results are combined and reconstructed to obtain the power load prediction data for the period to be predicted.

[0012] In a second aspect, an embodiment of the present application provides a short-term power load forecasting device, comprising:

[0013] An acquisition module is used to acquire historical load data of a historical period, historical meteorological data and historical time data corresponding to the historical load data, and future meteorological data and future time data of a period to be predicted;

[0014] A feature analysis module, used to perform feature analysis based on the historical load data, the historical meteorological data and the historical time data to obtain the features of the influencing factors of the historical load data;

[0015] An extraction module, for determining influencing factor data from the historical meteorological data, the historical time data, the future meteorological data and the future time data according to the influencing factor characteristics;

[0016] A modal decomposition module, used for performing modal decomposition on the historical load data to obtain multiple load components of different frequencies;

[0017] A prediction module, used to determine the prediction result corresponding to each load component according to a preset time series prediction model, influencing factor data and each load component;

[0018] The reconstruction module is used to merge and reconstruct various prediction results to obtain the power load prediction data for the time period to be predicted.

[0019] In a third aspect, an embodiment of the present application provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the steps of the method described in the first aspect or any possible implementation method of the first aspect are implemented.

[0020] In a fourth aspect, an embodiment of the present application provides a computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, it implements the steps of the method described in the first aspect or any possible implementation method of the first aspect.

[0021] Compared with the prior art, the embodiments of the present invention have the following beneficial effects:

[0022] The embodiment of the present application performs feature analysis through historical load data, historical meteorological data and historical time data, and can fully consider factors such as meteorology and time related to the load data to obtain the influencing factor characteristics of the historical load data, and then determine the influencing factor data from the historical meteorological data, historical time data, future meteorological data and future time data, which can retain the key features important for power load prediction, remove low-correlation features, and reduce the interference of low-correlation features in the prediction process, thereby improving the accuracy of the prediction. At the same time, it can also reduce data input and improve prediction efficiency; by performing modal decomposition on the historical load data, a plurality of load components with different frequencies can be obtained, and the nonlinearity and non-stationarity of the load data can be handled for subsequent accurate prediction; then, through the time series prediction model, the influencing factor data and each load component, each load component can be predicted separately, and the prediction result corresponding to each load component can be accurately obtained, avoiding the problem of inaccurate prediction caused by directly using historical load data for prediction; finally, by merging and reconstructing the various prediction results, the prediction data of the power load can be accurately obtained, and the prediction accuracy and performance can be improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0024] Figure 1 is a flow chart for implementing the short-term power load forecasting method provided in the embodiment of the present application;

[0025] Figure 2 It is a schematic diagram of the model structure of the short-term power load forecasting method provided in the embodiment of the present application;

[0026] Figure 3 is a schematic diagram of the structure of the TiDE model provided in the embodiment of the present application;

[0027] Figure 4 is a schematic diagram of the structure of the residual block provided in an embodiment of the present application;

[0028] Figure 5 is a schematic diagram of power load under different weather conditions provided by an embodiment of the present application;

[0029] Figure 6 is a histogram of the contribution values ​​of different types of data to the prediction provided by the embodiments of the present application;

[0030] Figure 7is a schematic diagram of component results of modal decomposition of a load sequence provided in an embodiment of the present application;

[0031] Figure 8 is a schematic diagram of a first power load prediction result under sunny conditions provided by an embodiment of the present application;

[0032] Fig. 9 This is a schematic diagram of a first power load prediction result under snowy conditions provided by an embodiment of the present application;

[0033] Fig.10 is a schematic diagram of a first power load prediction result under rainy conditions provided in an embodiment of the present application;

[0034] Fig.11 is a schematic diagram of a second power load forecast result under sunny conditions provided in an embodiment of the present application;

[0035] Fig.12 is a schematic diagram of a second power load prediction result under snowy conditions provided in an embodiment of the present application;

[0036] Fig.13 is a schematic diagram of a second power load prediction result under rainy conditions provided in an embodiment of the present application;

[0037] Fig.14 It is a structural schematic diagram of a short-term power load forecasting device provided in an embodiment of the present application;

[0038] Fig.15 It is a schematic diagram of an electronic device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0039] In the following description, specific details such as specific system structures, technologies, etc. are provided for the purpose of illustration rather than limitation, so as to provide a thorough understanding of the embodiments of the present application. However, it should be clear to those skilled in the art that the present application may also be implemented in other embodiments without these specific details. In other cases, detailed descriptions of well-known systems, devices, circuits, and methods are omitted to prevent unnecessary details from obstructing the description of the present application.

[0040] The inventor of the present application has found that in order to predict short-term power load, neural networks, machine learning models, and deep learning models are usually used for point prediction, which often fails to achieve the desired effect when predicting long time series. In addition, in rainy and snowy weather, the load prediction results will be significantly affected, making it difficult to accurately predict short-term power loads in complex environments. Therefore, it is necessary to consider a method for predicting short-term power loads in complex environments.

[0041] With the idea of ​​improving the accuracy of short-term power load forecasting in complex environments, in the implementation mode of the present application, feature analysis is performed on historical load data, historical meteorological data and historical time data to capture the key features in power load forecasting, and modal decomposition is performed on historical load data to obtain multiple load components of different frequencies. Each load component is predicted separately, and then the various prediction results are merged and reconstructed. This can handle the nonlinearity and non-stationarity of load data, achieve accurate prediction of short-term power load, and improve prediction accuracy and performance.

[0042] In order to make the purpose, technical solutions and advantages of the present application clearer, specific embodiments will be described below in conjunction with the accompanying drawings.

[0043] Figure 1 The implementation flow chart of the short-term power load forecasting method provided in the embodiment of the present application is detailed as follows:

[0044] Step 101, obtaining historical load data of a historical period, historical meteorological data and historical time data corresponding to the historical load data, and future meteorological data and future time data of a period to be predicted.

[0045] In this embodiment, the length of the historical period can be 1 year, 2 years, 3 years, etc., and the historical period is the period before the period to be predicted. The length of the period to be predicted can be any of the following: 4 hours, 8 hours, 12 hours, 24 hours, etc.

[0046] Meteorological data may include temperature, humidity, wind speed, rainfall, and snowfall, among which each item may be further divided, such as temperature may include maximum temperature, minimum temperature, and average temperature.

[0047] Time data may include types such as holidays, days of the week, and hours. For example, October 1, 2024 is a National Day holiday and falls on a Tuesday.

[0048] Step 102, performing feature analysis based on historical load data, historical meteorological data and historical time data to obtain the features of influencing factors of the historical load data.

[0049] In this embodiment, feature analysis can be performed on historical load data, historical meteorological data, and historical time data to determine the impact of each type of historical meteorological data and historical time data on power load forecasting, thereby selecting the type with a greater impact on load forecasting as the influencing factor feature.

[0050] Here, feature analysis can be performed using principal component analysis, correlation calculation, and machine learning models.

[0051] Optionally, before performing feature analysis on the historical power load data, meteorological data and date data to obtain load features and influencing factor features with high contribution values, missing value filling and normalization processing may be performed on the historical power load data, meteorological data and date data.

[0052] In this embodiment, in view of possible anomalies or data loss in the collection of power load and weather data, the raw data may be preprocessed to mitigate the adverse effects on prediction accuracy.

[0053] Here, the linear interpolation method can be used to fill the missing values. The specific formula can be:

[0054]

[0055] In the formula, y represents the missing value, x represents the horizontal coordinate corresponding to the missing value, and y 0 Represents the data of the collection point adjacent to the left side of the missing value, x 0 Represents y 0 The corresponding horizontal coordinate, y 1 Represents the data of the collection point adjacent to the right of the missing value, x 1 Represents y 1 The corresponding horizontal axis.

[0056] In order to solve the dimensionality difference, speed up the model convergence and improve the training accuracy, the data can be normalized. The specific formula is as follows:

[0057]

[0058] In the formula, y′ represents the normalized data, y represents the data before normalization, and Y min Represents the minimum value in the sample data, Y max Indicates the maximum value in the sample data.

[0059] Step 103, determining the influencing factor data from the historical meteorological data, the historical time data, the future meteorological data and the future time data according to the influencing factor characteristics.

[0060] In this embodiment, data corresponding to the characteristics of the influencing factors are selected from historical meteorological data, historical time data, future meteorological data and future time data as influencing factor data, so as to perform subsequent predictions, reduce the amount of data required for subsequent predictions, and avoid interference of low-relevance data in load predictions.

[0061] Step 104 , performing modal decomposition on the historical load data to obtain multiple load components of different frequencies.

[0062] In this embodiment, the load data is nonlinear and non-stationary, and it is difficult to make accurate predictions directly using historical load data. Therefore, the historical load data can be modally decomposed into multiple load components with different frequencies for subsequent predictions to reduce the prediction error caused by the nonlinearity and non-stationarity of the load while retaining the basic characteristics of the load.

[0063] Optionally, modal decomposition is performed on historical load data to obtain multiple load components of different frequencies, which may be performed by using variational nonlinear frequency modulation mode decomposition to perform modal decomposition on historical load data to obtain multiple load components of different frequencies.

[0064] In this embodiment, variational nonlinear chirp mode decomposition (VNCMD) can be used for modal decomposition, and the optimization problem in VNCMD is described and solved by the alternating multiplier method. The nonlinear chirp mode can be expressed by the following formula:

[0065]

[0066] Where g(t) represents the nonlinear frequency modulation function, A(t) represents the instantaneous amplitude of the signal, f(s) represents the instantaneous frequency of the signal, β represents the phase angle, and n(t) represents Gaussian white noise.

[0067] Step 105, determining the prediction result corresponding to each load component according to the preset time series prediction model, the influencing factor data and each load component.

[0068] In this embodiment, a time series prediction model may be used to predict each load component separately, wherein each load component may correspond to a model, and the hyperparameters of each model may be different, and may be specifically set or adjusted according to the corresponding load component.

[0069] Step 106, merging and reconstructing the various prediction results to obtain the power load prediction data for the time period to be predicted.

[0070] The embodiment of the present application performs feature analysis through historical load data, historical meteorological data and historical time data, and can fully consider factors such as meteorology and time related to the load data to obtain the influencing factor characteristics of the historical load data, and then determine the influencing factor data from the historical meteorological data, historical time data, future meteorological data and future time data, which can retain the key features important for power load prediction, remove low-correlation features, and reduce the interference of low-correlation features in the prediction process, thereby improving the accuracy of the prediction. At the same time, it can also reduce data input and improve prediction efficiency; by performing modal decomposition on the historical load data, a plurality of load components with different frequencies can be obtained, and the nonlinearity and non-stationarity of the load data can be processed for subsequent accurate prediction; then, through the time series prediction model, the influencing factor data and each load component, each load component can be predicted separately, and the prediction result corresponding to each load component can be accurately obtained, thereby avoiding the problem of inaccurate prediction caused by directly using historical load data for prediction; finally, by merging and reconstructing the various prediction results, the prediction data of the power load can be accurately obtained, and the prediction accuracy and performance can be improved.

[0071] In some embodiments, the historical weather data and the historical time data each include multiple types.

[0072] This embodiment performs feature analysis based on historical load data, historical meteorological data and historical time data to obtain the characteristics of influencing factors of historical load data. The historical load data, historical meteorological data and historical time data can be first input into a preset extreme gradient boosting model to perform feature analysis to obtain the contribution value of each type of extreme gradient boosting model output to the prediction; then, according to the contribution value corresponding to each type and the preset contribution threshold, the types of historical meteorological data and historical time data are screened respectively to obtain the characteristics of influencing factors of historical load data.

[0073] In this embodiment, the extreme gradient boosting model (EXtreme Gradient Boosting, XGBoost) can be used to perform feature analysis on historical meteorological data and historical time data to determine the contribution value of each type of historical meteorological data and historical time data to power load forecasting, that is, the correlation with power load in power load forecasting, so that historical meteorological data and historical time data can be screened, and types with high contribution values ​​can be retained to reduce the computational burden of subsequent predictions and improve prediction efficiency. In addition, the interference of data with low contribution values ​​can be reduced, the prediction accuracy can be improved, and the reliability of the final prediction results can be ensured.

[0074] Before feature analysis, the historical time data can be typed and reverse hot-encoded to convert it into a format that is easy for machine learning algorithms to interpret. Each type can be mapped to a separate binary vector to better capture category information.

[0075] The core of the XGBoost model is the decision tree, often referred to as a "weak learner". These decision trees are continuous in the XGBoost model. The generation of subsequent decision trees takes into account the predictions of previous decision trees, thereby taking into account the bias of previous trees during training. The attention mechanism ensures that samples that were predicted incorrectly receive greater attention in subsequent iterations. This design enhances XGBoost's ability to accurately capture data features, thereby improving prediction accuracy.

[0076] In addition, the training process of XGBoost starts from the objective function and iteratively builds a robust classifier. XGBoost uses a specialized gradient boosting tree algorithm to effectively improve the model accuracy and generalization ability through continuous iterations. XGBoost provides a feature importance evaluation method that can be used to identify features that have a significant impact on the classification results, promote model optimization and enhance prediction capabilities.

[0077] Optionally, the contribution threshold includes a meteorological contribution threshold and a time contribution threshold.

[0078] This embodiment screens the types of historical meteorological data and historical time data according to the contribution value corresponding to each type and the preset contribution threshold, and obtains the characteristics of influencing factors of historical load data. The types of historical meteorological data with contribution values ​​greater than the meteorological contribution threshold are first retained to obtain the meteorological influence type; then the types of historical time data with contribution values ​​greater than the time contribution threshold are retained to obtain the time influence type; finally, based on the meteorological influence type and the time influence type, the characteristics of influencing factors of historical load data are obtained.

[0079] In this embodiment, when screening, the difference between the contribution values ​​of different types of meteorological data and different types of time data may be large, and the meteorological contribution threshold and the time contribution threshold are set respectively. Here, the meteorological contribution threshold and the time contribution threshold can be the same or different, so as to accurately screen out the types with high correlation in contribution values ​​between meteorological data and time data.

[0080] In some embodiments, the influencing factor data includes dynamic influencing data and static influencing data; the time series prediction model includes multiple load prediction models, each load prediction model includes a feature projection module, a dense encoding module, a dense decoding module and a timing decoding module; each load component corresponds to a load prediction model.

[0081] The dynamic impact data mainly include the retained meteorological data and holiday data in the time data; the static impact data mainly include the retained time data except holiday data.

[0082] The load forecasting model can adopt the Time-series Dense Encoder (TiDE) model, the main goal of which is to accurately capture and predict the dynamic changes inherent in time series data. Figure 3 As shown in Figure 1, TiDE adopts the classic encoder-decoder architecture, and its entire model structure is composed entirely of multi-layer perceptrons (MLP). This design effectively solves the nonlinear problems of traditional linear models in modeling prediction and historical data, as well as the influence of external variables. The entire model can be divided into four parts: feature projection module, dense encoding module, dense decoding module, and time decoding module.

[0083] The TiDE model can overcome the challenges of capturing long-term dependencies and dealing with noise and uncertainty within the series. It cleverly combines the simplicity and effectiveness of linear models while effectively addressing the complexity of correlated variables and nonlinear dependencies. By deeply mining the temporal information embedded in time series data, TiDE can accurately identify future trends and patterns, thereby facilitating high-precision prediction of time series data.

[0084] The basic building block of this model is the residual block, which is Figure 4 As shown in the figure, it includes a dense layer plus a ReLU layer, a dense linear layer, and a summation layer normalization operation. Using this core component, other elements of TiDE can be built.

[0085] In this embodiment, the prediction result corresponding to each load component is determined according to the preset time series prediction model, the influencing factor data and each load component, which can be:

[0086] Step 301: for each load component, the dynamic impact data corresponding to the load component is input into the feature projection module of the corresponding load prediction model to obtain the low-dimensional impact data output by the feature projection module.

[0087] In this embodiment, the feature projection module can compress and project the input dynamic impact data, and reduce the dynamic impact data from a high-dimensional space to a low-dimensional space, so that the encoding module and the decoding module can process it more efficiently.

[0088] The feature projection module plays a specific role in managing dynamic impact data, which is closely related to time series forecasting and evolves over time, such as weather conditions or holidays. At this critical stage, the model uses advanced dimensionality reduction techniques to cleverly map dynamic impact data from high-dimensional space to low-dimensional space. This process can significantly simplify the amount of information that the model needs to process and reduce computational complexity, and can also accurately extract key features from dynamic impact data for prediction.

[0089] In essence, the feature projection module is an effective "compression" of the input data, enabling the encoder and decoder to quickly and accurately process this optimized information, thereby providing more precise information for time series prediction results, ensuring that the TiDE model has excellent prediction performance when faced with complex time series data.

[0090] Step 302: input the load component, and its corresponding static impact data and low-dimensional impact data into a dense coding module of a corresponding load prediction model to obtain coding information output by the dense coding module.

[0091] In this embodiment, the load components, static impact data and low-dimensional impact data may be combined in the time dimension and then input into a dense coding module for processing and encoding.

[0092] The dense encoding module plays a key role in converting load components, static influence data, and low-dimensional influence data (such as date and weather, as well as other external factors that affect the forecast) into internal representations. The encoding module deeply processes the input data through the constructed hierarchical network layers (i.e., MLP layers). Each layer can transform the data and extract insights from it, capturing the complex patterns and dependencies inherent in time series. The dense encoder can refine the raw input data into a compressed form of information, which is convenient for subsequent processing of the model and lays the foundation for further forecasting work.

[0093] Step 303: input the coding information into the dense decoding module of the corresponding load prediction model to obtain a decoding matrix output by the dense decoding module.

[0094] In this embodiment, the dense decoding module can decode the encoded information and convert it into a time series prediction value, i.e., a decoding matrix. The dense decoding module receives internal features from the dense encoder as input, and the dense decoding module initiates the construction of a prediction framework for future time points. A multi-layer perceptron layer is used to analyze the encoded information, identify potential patterns, and map these patterns to future related variables, thereby generating an accurate prediction of future time series values.

[0095] The dense decoding module can fully interpret the encoded data and transform it into specific and accurate predictions for future time points.

[0096] Step 304, input the decoding matrix and the low-dimensional impact data corresponding to the time period to be predicted into the time series decoding module of the corresponding load forecasting model, splice and decode according to the time dimension, and obtain the time series prediction data output by the time series decoding module.

[0097] In this embodiment, the temporal decoder can process the dense decoder output information and carefully process each time step to accurately transform the prediction generated by the dense decoder into the final time series output. By applying specific transformations at each time point, the temporal dependency of the prediction can be optimized, enhancing the model's ability to capture the complex temporal dynamics inherent in time series data.

[0098] The temporal decoder integrates the decoding process with the temporal dimension to produce accurate and progressively refined prediction results, thus providing strong support for time series forecasting tasks.

[0099] Step 305: Perform a global linear residual connection based on the load component and the time series prediction data to obtain a prediction result corresponding to the load component.

[0100] In this embodiment, during the prediction process of the load prediction model, a global linear residual connection can be set to improve the learning efficiency of the model and reduce the difficulty of training. Information can be directly transferred from the input load component to the subsequent layer, which helps to retain the information of the original input data, alleviate the problem of gradient disappearance, and ensure that the network can effectively learn and adapt to the training data, so as to accurately obtain the prediction results corresponding to the load component.

[0101] Optionally, for each load component, before inputting the dynamic impact data corresponding to the load component into the feature projection module of the corresponding load prediction model to obtain the low-dimensional impact data output by the feature projection module, a Bayesian optimization algorithm can be used to optimize the hyperparameters of the feature projection module, dense encoding module, dense decoding module and timing decoding module in the load prediction model corresponding to each load component to obtain the optimal hyperparameters of the load prediction model corresponding to each load component.

[0102] When using the TiDE model for prediction, the load components of different frequencies and the corresponding influencing factor data are input into each model for training and prediction. In order to ensure that each model uses the most appropriate hyperparameters, the Bayesian optimization module can be used to optimize the hyperparameters. Under the guidance of the optimized hyperparameters, each model effectively performs the prediction task and outputs the load prediction results of each component. Finally, the prediction results of each component are integrated and reconstructed to obtain the final load prediction results. The results are not only highly accurate, but also have important practical and guiding value.

[0103] Bayesian optimization is a black box optimization algorithm based on probability theory, which is good at using historical results to learn, optimize and adjust the parameters of complex functions. Its core concept is to guide the search process by establishing a probabilistic model of the objective function, thereby determining the parameter configuration that maximizes the value of the objective function.

[0104] In Bayesian optimization, the objective function is first assumed to follow a prior distribution, usually a Gaussian process model. Subsequently, in each iteration, the algorithm selects a point that is considered to be most likely to improve performance based on the probability model of the current objective function for evaluation. After the evaluation is completed, the new observations are incorporated into the model and the probability model is updated. This process is iterated until a preset number of iterations is reached or other stopping conditions are met.

[0105] Bayesian optimization can converge to a parameter configuration close to the optimal solution in fewer iterations by intelligently selecting the next evaluation point by leveraging historical observations. Second, Bayesian optimization can expertly handle complex objective functions, including multimodal and non-convex functions, making it very effective in a variety of practical applications. In addition, compared with traditional grid search or random search methods, Bayesian optimization explores the hyperparameter space more effectively and can identify the optimal hyperparameter configuration more quickly.

[0106] In some embodiments, the measured data from August 2022 to August 2023 in Province A is used for verification. The data set contains a total of 35,040 detailed records, and each data point is sampled at an interval of 15 minutes. The data set includes not only the power load data of the region, but also key characteristic parameters such as temperature, humidity, wind speed, rainfall, cloud cover, and snowfall, all of which are closely related to the power load.

[0107] In order to explore the differences in load characteristics under different weather conditions, the full dataset was divided into three subsets: sunny day data, rainy day data, and snowy day data. Then, we randomly selected 15-minute forecast curves for typical days under different weather conditions and displayed them on the Figure 5 middle.

[0108] from Figure 5 It can be clearly seen in the load curve fluctuation patterns on sunny, rainy and snowy days are very similar, consistent with typical electricity consumption habits. However, it is worth noting that during heavy snow or rain, load fluctuations become more obvious, with electricity consumption slightly higher than on sunny days. This is particularly evident during the peak load rise between points 60 and 80, providing a clear and convincing illustration of how weather conditions affect electricity load.

[0109] The specific verification process is divided into four parts. First, the XGBoost algorithm is used to perform in-depth correlation analysis on load and weather features to accurately identify features that are critical to prediction. Second, VNCMD decomposition is applied to the load time series to extract components of different frequencies, which are then used as input data for training the model. Next, the proposed model is comprehensively compared with LSTM, Transformer, Informer, and the model combined with VNCMD. The load forecasting accuracy under different weather conditions such as sunny, rainy, and snowy days is carefully evaluated, which fully demonstrates the superior performance of the proposed model. In addition, supplementary comparative experiments are carried out using real consistent data from specific regions, which confirms the good generalization ability of the proposed model. Finally, ablation experiments are carried out to explore the specific effects of VNCMD and Bayesian optimization on prediction accuracy, further verifying the effectiveness of the proposed model. It is worth noting that the feature correlation analysis, VNCMD decomposition, and model training parts of this study are all completed on the training set to ensure that no information leakage occurs.

[0110] First, historical load data, historical meteorological data, and historical time data are selected as the input features of the model, as shown in Table 1. The load data includes the load sequence starting from 4 hours before the forecast point, aiming to capture the short-term load change trend. At the same time, the meteorological data combines key meteorological indicators such as temperature, humidity, wind speed, rainfall, and snowfall to reflect the impact of weather factors on load fluctuations. In addition, the time data includes information about holidays, hours, and weeks to capture cyclic load patterns.

[0111] Daily, weekly, and holiday / non-holiday load data have obvious periodic characteristics, and time characteristics (such as hours, weeks, and holidays) are processed by one-hot encoding. One-hot encoding method is used to convert categorical variables into a format that is easy to interpret for machine learning algorithms. By mapping each type to an independent binary vector, this technique can better capture category information and improve the prediction accuracy of the model. Through this processing method, the time data is fully utilized, thereby further improving the prediction performance of the model.

[0112] Table 1 Type feature information table

[0113]

[0114] This example uses the XGBoost algorithm to calculate the contribution of each type and Figure 6 The corresponding histogram is given in . It is noteworthy that the forecast target shows an unusually high sensitivity to the historical load data, with particularly significant correlations. Specifically, the correlation of historical load data closer to the forecast target shows an upward trend, indicating that recent load data has a stronger predictive effect on future load forecasts.

[0115] In meteorological data, factors such as temperature, rainfall and snowfall are key factors affecting load. Especially in midsummer and cold winter, the increase of rainfall and snowfall often occurs simultaneously with drastic temperature fluctuations, which directly leads to the increase of power of high-power heating and cooling electrical equipment such as air conditioners. In addition, rainfall and snowfall affect people's travel patterns, which indirectly affects the power load.

[0116] In addition, time is a key indicator reflecting people's daily life and plays a vital role in load series forecasting. It covers time-related information such as holidays, hours, and weeks, which implies people's electricity consumption habits and patterns in different time periods.

[0117] By comprehensively analyzing the contributions of different types of data, they can be sorted from low to high according to their contribution values, and the types with small contribution values ​​can be eliminated. Reducing the computational burden of the model and improving computational efficiency can effectively improve prediction accuracy and ensure more accurate and reliable prediction results.

[0118] In order to further improve the accuracy of load series prediction, this paper applies the VNCMD method to decompose the load series. Figure 7 The component results of the modal decomposition of the load series are shown, where the y-axis represents the amplitude of each intrinsic mode function (IMF) component and the x-axis represents the time point. The VNCMD method can accurately capture the information of the load data, minimize the prediction error caused by the nonlinearity and non-stationarity of the load, and retain the basic characteristics of the load.

[0119] In order to further verify the effectiveness of the model proposed in this paper, it was compared with several baseline methods that have performed well in the field of time series prediction. Six models were selected, including three classic models: LSTM, Transformer and Informer, as well as their composite models combined with VNCMD. Due to its unique structural design, the LSTM model successfully solves the memory limitation problem of the RNN neural network, so that it can effectively model time series data. At the same time, the Transformer and Informer models introduce an attention mechanism to further optimize the performance, making them very suitable for processing long sequence data.

[0120] To ensure the fairness and objectivity of the experiments, the hyperparameters of the proposed model are carefully configured, as shown in Table 2. In addition, the parameters of other baseline comparison models are optimized to ensure that they achieve the best prediction performance, thus facilitating fair and convincing comparison conclusions.

[0121] Table 2 Model hyperparameters

[0122]

[0123]

[0124] In order to fully test the effectiveness and stability of the model, this paper uses the load series under different weather conditions in the test set as the test object. Specifically, a time period containing 96 data points (i.e., 24 hours) is selected for prediction. During the prediction process, the length of the input data is uniformly set to 96 points to ensure fair comparison between models under the same conditions. Subsequently, rolling predictions are performed using the various models mentioned above, and the predicted load data are recorded in detail.

[0125] Figure 8 , Fig. 9 and Fig.10 The test results of several days randomly selected from the test set under different weather conditions are shown for prediction comparison, where the ordinate represents the load and the abscissa represents the time, i.e. the sampling point. By comparing and analyzing these prediction results, it is obvious that the model proposed in this application shows good prediction performance in various weather types, with higher accuracy and stability compared to other baseline models. This fully proves that the proposed model performs well in dealing with complex and changing weather conditions.

[0126] from Figure 8 , Fig. 9 and Fig.10 It can be clearly seen that the model proposed in this application has significant advantages in terms of load forecast value fluctuations under various weather conditions (including sunny, snowy and rainy days). Compared with other baseline models, the power change between two adjacent sampling points in the model of this application is smoother, indicating better smoothness. Especially in the load ramp stage, such as 5h-10h and 15h-20h, the predicted value of the model of this application is closer to the actual power curve and has a smaller deviation. When the actual load reaches its peak, the model of this application can also be adjusted quickly and accurately, further verifying its high prediction accuracy.

[0127] In addition, under different weather conditions, the prediction accuracy of each model shows different characteristics. On sunny days, the prediction accuracy of all seven models is relatively high. However, under complex weather conditions such as rain and snow, the prediction sequence fluctuations of the other six models increased significantly, and their prediction performance was greatly affected. In contrast, the model proposed in this application maintained stable and accurate predictions, fully demonstrating its excellent stability. Compared with a single model, the prediction accuracy of the model combined with VNCMD has been significantly improved. This advantage is mainly due to the use of VNCMD decomposition and XGBoost feature selection methods, which effectively improve the model's anti-interference ability and prediction accuracy.

[0128] In addition, in order to more comprehensively evaluate the performance of each model, this embodiment selects a complete test set under different weather conditions for in-depth comparative analysis. As shown in Table 3, the quantitative evaluation results clearly demonstrate the excellent performance of the method proposed in this application. Compared with the other six methods, this method has significant advantages in prediction accuracy. Specifically, the prediction results of the proposed model not only accurately capture the periodicity and long-term change trends of the load data, but also maintain the highest accuracy in the test set under various weather conditions. Whether it is sunny, rainy or snowy, the model can consistently produce high-quality prediction results. In terms of evaluation indicators, the proposed model achieved the best levels of root mean square error (RMSE), mean absolute error (MAE) and mean absolute percentage error (MAPE).

[0129] Table 3 Prediction accuracy table

[0130]

[0131]

[0132] Among all the models involved in the comparison, the performance of the LSTM model is less satisfactory, which may be due to its limitations in handling long-term dependencies. In contrast, the Transformer and Informer models perform slightly better due to the introduction of the attention mechanism, which can better capture longer time dependencies. However, the model proposed in this application shows significant advantages in prediction performance. This is mainly due to the excellent prediction ability of the TiDE model, as well as the introduction of VNCMD decomposition and Bayesian optimization, which together improve the accuracy and stability of the model. Compared with the single model, the prediction accuracy of the model combined with VNCMD is improved, which is largely due to the benefits of VNCMD decomposition.

[0133] In the sunny weather forecast task, the RMSE error of the proposed model is 3.95 lower than that of the LSTM+VNCMD model, 2.81 lower than that of the Transformer+VNCMD model, and 2.22 lower than that of the Informer+VNCMD model. In the rainy weather forecast task, the RMSE error of the proposed model is 5.70 lower than that of the LSTM+VNCMD model, 4.79 lower than that of the Transformer+VNCMD model, and 3.54 lower than that of the Informer+VNCMD model. In the snowy weather forecast task, the RMSE error of the proposed model is 2.8 lower than that of the LSTM+VNCMD model, 0.83 lower than that of the Transformer+VNCMD model, and 0.59 lower than that of the Informer+VNCMD model.

[0134] In order to verify the universality and portability of the model proposed in this paper, the actual load data of Province B was selected for supplementary experiments. The data set includes load data and meteorological data, such as temperature, humidity, rainfall, and snowfall. The load forecast curves of the seven forecasting algorithms are shown in Figure 2. Fig.11 , Fig.12 and Fig.13 As shown, the error evaluation indicators of the seven prediction methods are shown in Table 4. It can be clearly seen that the prediction result of the proposed model is the best, its curve is closest to the true value, and it shows the highest prediction accuracy.

[0135] Table 4 Prediction accuracy table

[0136]

[0137]

[0138] A series of ablation experiments were also conducted in this example to explore the specific effects of VNCMD and Bayesian optimization on the prediction performance of the TiDE model. The TiDE model was used as a baseline comparison model, and a 96-point short-term prediction experiment was conducted in combination with VNCMD and Bayesian optimization. The detailed recording and analysis of the model prediction errors revealed the following conclusions. The prediction accuracy of different models is shown in Table 5.

[0139] First, the introduction of VNCMD significantly improves the prediction performance of the TiDE model. The TiDE+VNCMD model uses VNCMD decomposition technology to effectively enhance the model's ability to handle multi-frequency features, thereby improving prediction accuracy. This improvement enables the model to better capture the inherent patterns and trends of load data, thereby achieving more accurate predictions.

[0140] Secondly, the application of Bayesian optimization also had a positive impact on the prediction performance of the TiDE model. The TiDE+Bayesian model further improved the prediction accuracy by optimizing hyperparameters. The Bayesian optimization method effectively searches for the optimal solution in the parameter space, improving the generalization ability and stability of the model.

[0141] It is particularly noteworthy that when VNCMD and Bayesian optimization are combined with the TiDE model, a combined model with the best prediction performance can be obtained. The experimental results fully demonstrate the effectiveness of VNCMD and Bayesian optimization in improving prediction accuracy. By combining the two, a prediction model that can handle complex multi-frequency characteristics and has excellent hyperparameter configuration is successfully constructed, achieving high-precision prediction of load data.

[0142] Table 5 Prediction accuracy table

[0143]

[0144]

[0145] It can be seen that the short-term power load forecasting method provided in this application has higher prediction accuracy than LSTM, Transformer, Informer and their combination with VNCMD, and performs well in extreme rainy and snowy weather, showing good stability and versatility.

[0146] It should be understood that the size of the serial numbers of the steps in the above embodiments does not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.

[0147] The following is an embodiment of the device of the present application. For details not described in detail, please refer to the corresponding method embodiment described above.

[0148] Fig.14 The following is a schematic diagram of the structure of a short-term power load forecasting device provided in an embodiment of the present application. For ease of explanation, only the parts related to the embodiment of the present application are shown, which are described in detail as follows:

[0149] like Fig.14 As shown, the short-term power load forecasting device 140 includes:

[0150] The acquisition module 1401 is used to acquire historical load data of a historical period, historical meteorological data and historical time data corresponding to the historical load data, and future meteorological data and future time data of a period to be predicted;

[0151] The feature analysis module 1402 is used to perform feature analysis based on historical load data, historical meteorological data and historical time data to obtain the features of influencing factors of the historical load data;

[0152] Extraction module 1403, used to determine influencing factor data from historical meteorological data, historical time data, future meteorological data and future time data according to the influencing factor characteristics;

[0153] The modal decomposition module 1404 is used to perform modal decomposition on the historical load data to obtain multiple load components of different frequencies;

[0154] The prediction module 1405 is used to determine the prediction result corresponding to each load component according to the preset time series prediction model, the influencing factor data and each load component;

[0155] The reconstruction module 1406 is used to merge and reconstruct various prediction results to obtain the power load prediction data for the time period to be predicted.

[0156] In a possible implementation, the historical meteorological data and the historical time data include multiple types respectively; the extraction module 1403 is specifically used for:

[0157] Input historical load data, historical meteorological data and historical time data into the preset extreme gradient boosting model, perform feature analysis, and obtain the contribution value of each type of extreme gradient boosting model output to the prediction;

[0158] According to the contribution value corresponding to each type and the preset contribution threshold, the types of historical meteorological data and historical time data are screened respectively to obtain the characteristics of influencing factors of historical load data.

[0159] In a possible implementation, the contribution threshold includes a meteorological contribution threshold and a time contribution threshold;

[0160] The extraction module 1403 is specifically used for:

[0161] The types of historical meteorological data whose contribution values ​​are greater than the meteorological contribution threshold are retained to obtain the meteorological impact type;

[0162] The types whose contribution values ​​in the historical time data are greater than the time contribution threshold are retained to obtain the time impact type;

[0163] Based on the meteorological impact type and time impact type, the influencing factor characteristics of historical load data are obtained.

[0164] Fig.15 Schematic diagram of an electronic device provided in an embodiment of the present application. Fig.15 As shown, the electronic device 150 of this embodiment includes: a processor 151, a memory 152, and a computer program 153 stored in the memory 152 and executable on the processor 151. When the processor 151 executes the computer program 153, the steps in the above-mentioned various short-term power load forecasting method embodiments are implemented, such as Figure 1 Alternatively, the processor 151 executes the computer program 153 to implement the functions of each module in the above-mentioned device embodiments, such as Fig.14 Functions of modules 1401 to 1406 are shown.

[0165] Exemplarily, the computer program 153 may be divided into one or more modules / units, one or more modules / units are stored in the memory 152 and executed by the processor 151 to complete the present application. One or more modules / units may be a series of computer program instruction segments that can complete specific functions, and the instruction segments are used to describe the execution process of the computer program 153 in the electronic device 150. For example, the computer program 153 may be divided into Fig.14 Modules 1401 to 1406 are shown.

[0166] The electronic device 150 may include, but is not limited to, a processor 151 and a memory 152. Those skilled in the art will appreciate that Fig.15 It is only an example of the electronic device 150 and does not constitute a limitation of the electronic device 150. It may include more or fewer components than shown in the figure, or a combination of certain components, or different components. For example, the electronic device may also include input and output devices, network access devices, buses, etc.

[0167] The processor 151 may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or any conventional processor, etc.

[0168] The memory 152 may be an internal storage unit of the electronic device 150, such as a hard disk or memory of the electronic device 150. The memory 152 may also be an external storage device of the electronic device 150, such as a plug-in hard disk, a smart memory card (Smart Media Card, SMC), a secure digital (Secure Digital, SD) card, a flash card (FlashCard), etc. equipped on the electronic device 150. Further, the memory 152 may also include both an internal storage unit of the electronic device 150 and an external storage device. The memory 152 is used to store computer programs and other programs and data required by the electronic device. The memory 152 may also be used to temporarily store data that has been output or is to be output.

[0169] The technicians in the relevant field can clearly understand that for the convenience and simplicity of description, only the division of the above-mentioned functional units and modules is used as an example for illustration. In practical applications, the above-mentioned function allocation can be completed by different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiment can be integrated in a processing unit, or each unit can exist physically separately, or two or more units can be integrated in one unit. The above-mentioned integrated unit can be implemented in the form of hardware or in the form of software functional units. In addition, the specific names of the functional units and modules are only for the convenience of distinguishing each other, and are not used to limit the scope of protection of this application. The specific working process of the units and modules in the above-mentioned system can refer to the corresponding process in the aforementioned method embodiment, which will not be repeated here.

[0170] In the above embodiments, the description of each embodiment has its own emphasis. For parts that are not described or recorded in detail in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0171] Those of ordinary skill in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application.

[0172] In the embodiments provided in the present application, it should be understood that the disclosed devices / electronic devices and methods can be implemented in other ways. For example, the device / electronic device embodiments described above are merely schematic. For example, the division of modules or units is only a logical function division. There may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.

[0173] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0174] In addition, each functional unit in each embodiment of the present application may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of software functional units.

[0175] If the integrated module / unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the present application implements all or part of the processes in the above-mentioned embodiment method, and can also be completed by instructing the relevant hardware through a computer program. The computer program can be stored in a computer-readable storage medium, and the computer program can implement the steps of the above-mentioned various method embodiments when executed by the processor. Among them, the computer program includes computer program code, and the computer program code can be in source code form, object code form, executable file or some intermediate form. Computer-readable media may include: any entity or device capable of carrying computer program code, recording medium, U disk, mobile hard disk, disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electric carrier signal, telecommunication signal and software distribution medium, etc.

[0176] The above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. These modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application, and should all be included in the protection scope of the present application.

Claims

1. A short-term power load forecasting method, characterized in that: include: Obtain historical load data for a historical period, historical meteorological data and historical time data corresponding to the historical load data, and future meteorological data and future time data for a period to be predicted; Performing feature analysis based on the historical load data, the historical meteorological data, and the historical time data to obtain the features of influencing factors of the historical load data; Determining influencing factor data from the historical meteorological data, the historical time data, the future meteorological data and the future time data according to the influencing factor characteristics; Performing modal decomposition on the historical load data to obtain multiple load components of different frequencies; Determine the prediction result corresponding to each load component according to the preset time series prediction model, influencing factor data and each load component; The various prediction results are combined and reconstructed to obtain the power load prediction data for the period to be predicted.

2. The short-term power load forecasting method according to claim 1, characterized in that: The historical meteorological data and the historical time data include multiple types respectively; The performing of feature analysis based on the historical load data, the historical meteorological data and the historical time data to obtain the features of the influencing factors of the historical load data includes: Inputting the historical load data, the historical meteorological data and the historical time data into a preset extreme gradient boosting model, performing feature analysis, and obtaining a contribution value of each type of output of the extreme gradient boosting model to the prediction; According to the contribution value corresponding to each type and the preset contribution threshold, the types of the historical meteorological data and the historical time data are screened respectively to obtain the influencing factor characteristics of the historical load data.

3. The short-term power load forecasting method according to claim 2, characterized in that: The contribution threshold includes a meteorological contribution threshold and a time contribution threshold; According to the contribution value corresponding to each type and the preset contribution threshold, the types of the historical meteorological data and the historical time data are screened respectively to obtain the influencing factor characteristics of the historical load data, including: The types of the historical meteorological data whose contribution values ​​are greater than the meteorological contribution threshold are retained to obtain the meteorological impact type; Retain the types of the historical time data whose contribution values ​​are greater than the time contribution threshold to obtain the time impact type; Based on the meteorological impact type and the time impact type, the influencing factor characteristics of the historical load data are obtained.

4. The short-term power load forecasting method according to claim 1, characterized in that: The influencing factor data includes dynamic influencing data and static influencing data; the time series prediction model includes multiple load prediction models, each load prediction model includes a feature projection module, a dense encoding module, a dense decoding module and a time series decoding module; each load component corresponds to a load prediction model; Determining the prediction result corresponding to each load component according to the preset time series prediction model, the influencing factor data and each load component includes: For each load component, the dynamic impact data corresponding to the load component is input into the feature projection module of the corresponding load prediction model to obtain the low-dimensional impact data output by the feature projection module; Input the load component, and its corresponding static impact data and low-dimensional impact data into a dense coding module of a corresponding load forecasting model to obtain coding information output by the dense coding module; Inputting the encoded information into a dense decoding module of a corresponding load forecasting model to obtain a decoding matrix output by the dense decoding module; Input the decoding matrix and the low-dimensional impact data corresponding to the time period to be predicted into the time series decoding module of the corresponding load forecasting model, and perform splicing and decoding according to the time dimension to obtain the time series forecasting data output by the time series decoding module; A global linear residual connection is performed based on the load component and the time series prediction data to obtain a prediction result corresponding to the load component.

5. The short-term power load forecasting method according to claim 4, characterized in that: Before inputting the dynamic impact data corresponding to each load component into the feature projection module of the corresponding load prediction model to obtain the low-dimensional impact data output by the feature projection module, the method further includes: The Bayesian optimization algorithm is used to optimize the hyperparameters of the feature projection module, dense encoding module, dense decoding module and temporal decoding module in the load forecasting model corresponding to each load component, and the optimal hyperparameters of the load forecasting model corresponding to each load component are obtained.

6. The short-term power load forecasting method according to claim 1, characterized in that: The modal decomposition of the historical load data is performed to obtain multiple load components of different frequencies, including: The historical load data is modally decomposed by using variational nonlinear frequency modulation mode decomposition to obtain multiple load components with different frequencies.

7. The short-term power load forecasting method according to any one of claims 1 to 6, characterized in that: Before performing feature analysis on the historical power load data, the meteorological data and the date data to obtain load features and influencing factor features with high contribution values, the method further includes: The historical power load data, the meteorological data and the date data are subjected to missing value filling and normalization processing.

8. A short-term power load forecasting device, characterized in that: include: An acquisition module is used to acquire historical load data of a historical period, historical meteorological data and historical time data corresponding to the historical load data, and future meteorological data and future time data of a period to be predicted; A feature analysis module, used to perform feature analysis based on the historical load data, the historical meteorological data and the historical time data to obtain the features of the influencing factors of the historical load data; An extraction module, for determining influencing factor data from the historical meteorological data, the historical time data, the future meteorological data and the future time data according to the influencing factor characteristics; A modal decomposition module, used for performing modal decomposition on the historical load data to obtain multiple load components of different frequencies; A prediction module, used to determine the prediction result corresponding to each load component according to a preset time series prediction model, influencing factor data and each load component; The reconstruction module is used to merge and reconstruct various prediction results to obtain the power load prediction data for the time period to be predicted.

9. An electronic device comprising a memory and a processor, wherein the memory is used to store a computer program, and the processor is used to call and run the computer program stored in the memory, characterized in that: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 7 are implemented.

Images