Power load prediction method and system based on adaptive time series decomposition and electronic device

Abstract

The application discloses a power load prediction method and system based on adaptive time series decomposition and electronic equipment, and relates to the technical field of power system load prediction. The method obtains historical power load data and pre-processes the same, constructs an adaptive time series decomposition model to decompose the pre-processed load data into a plurality of subsequence components, wherein the number of decomposition layers is adaptively determined according to the characteristics of the load data. Stationarity test is performed on each subsequence component, and corresponding prediction models are matched for each subsequence component according to the test results for prediction. The prediction results of the subsequence components are integrated and reconstructed to obtain a final power load prediction result. The application can adaptively process load data with different characteristics, significantly improve the accuracy and stability of power load prediction, and is suitable for power load prediction scenarios of different time scales.

Description

[0001] This invention belongs to the field of power system load forecasting technology, and particularly relates to a power load forecasting method, system and electronic equipment based on adaptive time series decomposition. Background Technology

[0002] Electricity load forecasting is a crucial foundation for power system planning, operation, and dispatch. Accurate load forecasting is of paramount importance for ensuring the safe and stable operation of the power grid and improving energy efficiency. With the deepening of smart grid construction and power market reforms, the accuracy requirements for electricity load forecasting are constantly increasing, and the scale and complexity of load data are also significantly growing.

[0003] Existing power load forecasting methods mainly fall into two categories: traditional statistical forecasting methods and artificial intelligence forecasting methods. Traditional statistical forecasting methods, such as autoregressive integral moving average models and exponential smoothing, have good predictive performance for load sequences with obvious linear characteristics. However, power load data typically exhibits strong nonlinear and non-stationary characteristics, making it difficult for traditional statistical methods to fully capture the complex variation patterns within the load data. Artificial intelligence forecasting methods, such as neural networks and support vector machines, while possessing strong nonlinear fitting capabilities, often struggle to maintain the same prediction accuracy for different components when dealing with composite load signals containing multiple frequency components and varying variation patterns.

[0004] Temporal decomposition techniques offer an effective solution to the aforementioned problems. By decomposing the original load sequence into several components with different characteristics and then modeling and predicting each component separately, the complexity of the original sequence and the difficulty of prediction can be effectively reduced. However, existing temporal decomposition prediction methods mostly employ fixed decomposition levels and fixed prediction model matching strategies, failing to adaptively adjust according to the actual characteristics of the load data, resulting in significant fluctuations in prediction performance under different load data characteristics. Furthermore, existing methods lack quantitative evaluation of the decomposition effect and adaptive optimization mechanisms during the decomposition process, making it difficult to guarantee the rationality of the decomposition results and the accuracy of prediction. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a power load forecasting method, system, and electronic device based on adaptive time-series decomposition, aiming to solve the technical problem of insufficient forecasting accuracy caused by fixed decomposition strategies and the inability to adaptively match load data characteristics in existing power load forecasting methods. Technical solution

[0006] To achieve the above objectives, the present invention provides the following technical solution: a power load forecasting method based on adaptive time-series decomposition, comprising the following steps: S1. Obtain historical power load data and preprocess the historical power load data; S2. Construct an adaptive time series decomposition model to decompose the preprocessed load data into several subsequence components. The number of decomposition layers of the adaptive time series decomposition model is adaptively determined based on the characteristic parameters of the load data. S3. Perform stationarity tests on each subsequence component, match the corresponding prediction model to each subsequence component based on the test results, and use the matched prediction model to predict each subsequence component to obtain the prediction results of each subsequence component. S4. Integrate and reconstruct the prediction results of each subsequence component to obtain the power load prediction result.

[0007] As a preferred embodiment of the present invention, the preprocessing in step S1 includes: Historical power load data is processed for missing values, outlier detection and correction, and data normalization. The missing value processing adopts linear interpolation or neighbor-value filling methods, and the outlier detection is based on the statistical distribution characteristics or rate of change threshold of the load data.

[0008] As a preferred embodiment of the present invention, the adaptive time-series decomposition model is constructed based on the variational mode decomposition algorithm, and the number of decomposition layers is adaptively determined according to the center frequency distribution characteristics of the load data and the evaluation index of mode aliasing, including: The load data is decomposed layer by layer according to the preset decomposition layer range. The ratio of the modal center frequency distribution under each decomposition layer is calculated. When the ratio of the center frequencies of adjacent modes is less than the preset frequency threshold, the decomposition layer is determined as the optimal decomposition layer.

[0009] As a preferred embodiment of the present invention, the adaptive time series decomposition model is constructed based on the ensemble empirical mode decomposition algorithm, and the number of decomposition layers is adaptively determined according to the extreme point distribution density and energy loss rate of the load data; the extreme point distribution density is calculated according to the formula =, where M is the total number of local extreme points in the load time series and N is the sequence length; The energy loss rate is calculated according to the formula =, where is the original load signal energy, is the energy of the Kth modal component, and K is the decomposition layer number; K is increased layer by layer and the corresponding value is calculated. When the decrease rate of K is less than the preset threshold for the first time, the current K is determined as the optimal decomposition layer number.

[0010] In a preferred embodiment of the present invention, step S3 involves performing a stationarity test on each subsequence component and matching a corresponding prediction model to each subsequence component based on the test results, including: The unit root test method was used to test the stationarity of each subsequence component. Subsequence components that pass the stationarity test are identified as stationary components, and subsequence components that fail the stationarity test are identified as non-stationary components. The stationary component is matched with a linear prediction model, and the non-stationary component is matched with a nonlinear prediction model.

[0011] As a preferred embodiment of the present invention, the linear prediction model is an autoregressive integral moving average model or an autoregressive model, and the nonlinear prediction model is a long short-term memory network model, a gated recurrent unit model or a support vector regression model.

[0012] In a preferred embodiment of the present invention, step S4 involves integrating and reconstructing the prediction results of each sub-sequence component, including: The prediction results of each subsequence component are linearly superimposed to obtain the initial reconstruction result; The initial reconstruction result is subjected to error correction processing, which is based on the statistical characteristics of historical prediction errors or the predicted values ​​of the error sequence.

[0013] As a preferred embodiment of the present invention, the characteristic parameters include at least one of the following: the fluctuation frequency of the load data, the range of amplitude variation, and the trend strength.

[0014] Furthermore, this invention also proposes a power load forecasting system based on adaptive time-series decomposition, comprising: The data acquisition module is used to acquire historical power load data and preprocess the historical power load data; The adaptive decomposition module is used to construct an adaptive time-series decomposition model, which decomposes the preprocessed load data into several sub-series components. The number of decomposition layers of the adaptive time-series decomposition model is adaptively determined based on the characteristic parameters of the load data. The component prediction module is used to perform stationarity tests on each subsequence component, match the corresponding prediction model for each subsequence component based on the test results, and use the matched prediction model to predict each subsequence component to obtain the prediction result of each subsequence component. The reconstruction module is used to integrate and reconstruct the prediction results of each subsequence component to obtain the power load prediction result.

[0015] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the method described above.

[0016] This invention constructs an adaptive time-series decomposition model, allowing the number of decomposition levels to be adaptively determined based on the characteristic parameters of the load data. This avoids the process decomposition or under-decomposition problems caused by a fixed number of decomposition levels, and improves the matching degree between the decomposition results and the inherent characteristics of the load data. By performing stationarity tests on each subsequence component after decomposition and matching the corresponding prediction model accordingly, the most suitable prediction strategy can be adopted for components with different characteristics, fully exploring the variation patterns of each component and effectively improving prediction accuracy and model generalization ability. Validated with actual power load data, the prediction accuracy of this invention is significantly better than existing prediction methods with fixed decomposition strategies under different seasons and time scales, demonstrating broad application value.

[0017] Compared with existing technologies, this invention provides a power load forecasting method, system, and electronic equipment based on adaptive time-series decomposition, which has the following advantages: I. This invention constructs an adaptive time-series decomposition model, allowing the number of decomposition levels to be adaptively determined based on the characteristic parameters of the load data. It automatically selects the optimal number of decomposition levels according to the fluctuation frequency, amplitude range, and trend strength of the actual load data, effectively avoiding the process decomposition or under-decomposition problems caused by a fixed number of decomposition levels. This improves the matching degree between the decomposition results and the inherent characteristics of the load data, thereby significantly enhancing the accuracy and stability of subsequent predictions. Furthermore, the prediction method provided by this invention is applicable to power load prediction scenarios at different time scales, possessing good generalization ability and adaptability. It can be widely applied to various application scenarios such as day-ahead load prediction, short-term load prediction, and medium- and long-term load prediction, providing strong technical support for power system operation scheduling and planning decisions. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the method flow of the present invention; Figure 2 This is a schematic diagram of the system module structure of the present invention. Detailed Implementation

[0019] To better understand the purpose, structure, and function of this invention, the following will further describe the power load forecasting method, system, and electronic equipment based on adaptive time-series decomposition based on specific embodiments with reference to the accompanying drawings. It should be understood that the power load forecasting method described in the embodiments of this invention can be applied to various application scenarios requiring load forecasting, such as power system dispatching agencies, power generation companies, grid operating units, and power market trading centers. The historical power load data in the embodiments of this invention can originate from grid dispatch automation systems, data acquisition and monitoring control systems, or electricity consumption information acquisition systems, etc. Example

[0020] This embodiment provides a power load forecasting method based on adaptive time-series decomposition, which includes the following steps: Step S1: Obtain historical power load data and preprocess the historical power load data.

[0021] Specifically, the historical power load data comes from the power system's data acquisition and monitoring control system. The data type is time series data, with a sampling interval of 15 minutes, meaning that there are 96 load data points per day. The acquired raw load data contains some missing data points and outliers, which require preprocessing.

[0022] The preprocessing process includes three sub-steps: missing value handling, outlier detection and correction, and data normalization.

[0023] For handling missing values, when the number of consecutively missing data points is less than or equal to 3, linear interpolation is used to fill them, that is, the load values ​​of adjacent time points before and after the missing point are used for linear interpolation calculation; when the number of consecutively missing data points is greater than 3, the nearest neighbor value is used to fill them, that is, the load value of the same time point on the day before the missing point is used for filling.

[0024] For outlier detection and correction, a detection method based on statistical distribution characteristics is adopted, specifically: the mean and standard deviation of historical load data are calculated, and load data points that deviate from the mean by more than three times the standard deviation are initially identified as outliers; at the same time, a verification is performed in conjunction with the rate of change threshold, and points where the load change rate at adjacent time points exceeds the preset rate of change threshold are identified as outliers; for load data points identified as outliers, median filtering or the mean of adjacent time points is used for correction; for data normalization, the max-min normalization method is used to map the load data to the interval [0,1]; after preprocessing, the normalized load data is divided into training set and test set according to a certain ratio.

[0025] In this embodiment, the first 80% of the data is used as the training set for model building and parameter estimation, and the last 20% of the data is used as the test set for prediction performance evaluation.

[0026] Step S2: Construct an adaptive time series decomposition model to decompose the preprocessed load data into several subsequence components. The number of decomposition layers of the adaptive time series decomposition model is adaptively determined based on the characteristic parameters of the load data.

[0027] Specifically, this embodiment uses variational mode decomposition algorithm to construct an adaptive time-series decomposition model. Variational mode decomposition algorithm is a non-recursive signal decomposition method that decomposes the original signal into several band-limited mode components around their respective center frequencies by iteratively solving variational problems.

[0028] In variational mode decomposition, the determination of the number of decomposition layers K is a key factor affecting the decomposition effect. If the value of K is set too small, different frequency components in the original signal will be mixed in the same mode, resulting in under-decomposition. If the value of K is set too large, the same frequency component will be split into multiple modes, resulting in over-decomposition. Both of these situations will reduce the accuracy of subsequent predictions.

[0029] In this embodiment, the number of decomposition layers is adaptively determined based on the center frequency distribution characteristics of the load data and the evaluation index of modal aliasing degree. The specific process is as follows: The preset search range for the number of decomposition layers K is [2, 10]. That is, variational mode decomposition of the load data is performed layer by layer from K=2 to K=10. For each K value, K modal components are obtained after decomposition, and the center frequency of each modal component is calculated. When the number of decomposition layers K is appropriate, the center frequencies of each modal component exhibit a uniform distribution in the frequency domain. When the K value exceeds the optimal number of decomposition layers, the center frequencies of adjacent modes may become too close or even overlap.

[0030] Meanwhile, to enhance the reliability of the adaptive determination results, a modal aliasing degree evaluation index is introduced for comprehensive judgment. The modal aliasing degree is evaluated by calculating the correlation coefficient between each modal component. When the correlation coefficient between any two modal components exceeds a preset correlation threshold, modal aliasing is determined to exist at that decomposition level. Combining the center frequency distribution ratio and the modal aliasing degree, the optimal decomposition level is determined based on both satisfying preset conditions.

[0031] In this embodiment, the characteristic parameters of the load data include the fluctuation frequency, amplitude variation range, and trend strength. The fluctuation frequency is obtained through spectral analysis of the original load data, the amplitude variation range is obtained by calculating the difference between the maximum and minimum values ​​of the load data, and the trend strength is obtained by calculating the linear trend slope of the load data. These characteristic parameters are implicitly reflected in the calculation process of the center frequency distribution ratio, which directly reflects the composition of different frequency components in the load data.

[0032] Step S3: Perform stationarity tests on each subsequence component, match the corresponding prediction model to each subsequence component based on the test results, and use the matched prediction model to predict each subsequence component to obtain the prediction results of each subsequence component.

[0033] Specifically, this embodiment uses the augmented Dickey-Fuller test, a unit root test method, to perform stationarity tests on the four subsequence components obtained in step S1. The basic principle of the augmented Dickey-Fuller test is as follows: for a given time series, a regression equation containing lag terms is constructed. The presence of a unit root is determined by testing whether the coefficient of the unit root term in the regression equation is significantly different from zero. If the test statistic is less than the critical value, the null hypothesis of the existence of a unit root is rejected, and the series is determined to be stationary; otherwise, the series is determined to be non-stationary.

[0034] The test results are as follows: The P-value for the IMF1 component is 0.002, which is less than the significance level of 0.05, and it passes the stationarity test, thus being determined to be a stationary component; the P-value for the IMF2 component is 0.015, and it passes the stationarity test, thus being determined to be a stationary component; the P-value for the IMF3 component is 0.078, and it fails the stationarity test, thus being determined to be a non-stationary component; the P-value for the IMF4 component is 0.321, and it fails the stationarity test, thus being determined to be a non-stationary component.

[0035] Based on the test results, in order to match the linear prediction model of IMF1 and IMF2 that have passed the stationarity test, the linear prediction model adopted in this embodiment is the autoregressive integral moving average model; for the IMF1 component, the model order is determined to be ARIMA(2,0,1) through autocorrelation function and partial autocorrelation function analysis; for the IMF2 component, the model order is determined to be ARIMA(3,0,2).

[0036] To match the nonlinear prediction models for IMF3 and IMF4, which failed the stationarity test, this embodiment employs a Long Short-Term Memory (LSTM) network model. The LSTM model is a special recurrent neural network structure that effectively solves the gradient vanishing and gradient exploding problems of traditional recurrent neural networks when processing long sequences by introducing a gating mechanism. For the IMF3 component, a network structure with two LSTM hidden layers is constructed, with 64 hidden units per layer, and the output layer is a fully connected layer. For the IMF4 component, the same network structure is constructed, but the number of hidden units is adjusted to 32. The Adam optimizer is used for training each LSTM model, the mean squared error loss function is used, the number of training epochs is set to 200, and the batch size is set to 32.

[0037] The matched prediction model described above is used to predict the four subsequence components IMF1, IMF2, IMF3, and IMF4, respectively. The prediction step size is set according to actual needs. In this embodiment, the prediction step size is set to 96 steps, that is, to predict the load data for the next day. The prediction results of each component are denoted as , , and .

[0038] Step S4: Integrate and reconstruct the prediction results of each subsequence component to obtain the power load prediction result.

[0039] In the initial reconstruction sub-step, the prediction results of each sub-sequence component are linearly superimposed to obtain the initial reconstruction result. The linear superposition method is as follows: the predicted values ​​of , , and at the corresponding times are added together to obtain the initial reconstruction load sequence = ++, where t is the prediction time. Example

[0040] This embodiment provides a power load forecasting method based on adaptive time series decomposition. The main difference between this embodiment and Embodiment 1 lies in the construction method of the adaptive time series decomposition model.

[0041] In this embodiment, the adaptive time series decomposition model is constructed based on the ensemble empirical mode decomposition algorithm, which is an improved algorithm developed on the basis of empirical mode decomposition. By adding white noise to the original signal and performing multiple empirical mode decompositions, the statistical characteristics of white noise are used to overcome the mode aliasing problem in empirical mode decomposition.

[0042] The number of decomposition layers is adaptively determined based on the extreme point distribution density and energy loss rate. Let the load time series length be N, and the total number of local maxima and minima in the detection sequence be M. Then the extreme point distribution density is = . The original load signal energy is = , and the sum of the energy of each modal component when the decomposition layer number is K is , where = , and the energy loss is expressed as follows: = . Starting from K=1, the layers are increased sequentially, and the value corresponding to each K value is calculated, along with the decrease in energy between adjacent layers is calculated. When the value is first less than a preset threshold (taken as == 0.5% in this embodiment), the current K is determined as the optimal decomposition layer number. In this embodiment, through the above adaptive calculation, the optimal decomposition layer number is determined to be 5, and then ensemble empirical mode decomposition is performed on the load data according to this number of layers.

[0043] The remaining steps are the same as in Example 1, and will not be repeated here. Example

[0044] This embodiment provides a power load forecasting method based on adaptive time series decomposition. The main difference between this embodiment and Embodiment 1 lies in the selection and matching strategy of the component forecasting model.

[0045] In this embodiment, the method for testing the stationarity of each subsequence component is the same as in Embodiment 1, using the augmented Dickey-Fuller test. However, for different stationarity test results, this embodiment provides more prediction model selection options.

[0046] Specifically, for stationary components that pass the stationarity test, in addition to using the autoregressive integral moving average (ARIMA) model, an autoregressive (AR) model can also be used. An autoregressive model is a model that describes the linear relationship between the current value and past values, and its model order is determined by the Akaike information criterion or the Bayesian information criterion. In this embodiment, the stationary component IMF1 is matched with the autoregressive model AR(4), and IMF2 is matched with the autoregressive model AR(6).

[0047] For non-stationary components that fail the stationarity test, in addition to using a Long Short-Term Memory (LSTM) network model, a Gated Recurrent Unit (GRU) model or a Support Vector Regression (SVR) model can also be used. The GRU model is another improved recurrent neural network structure, which has fewer gating structures and lower computational complexity compared to LSTM networks. In this embodiment, the GRU model is used to predict the non-stationary component IMF3, and the SVR model is used to predict the non-stationary component IMF4.

[0048] Support vector regression (SVR) is based on the principle of structural risk minimization. It maps input data to a high-dimensional feature space through a kernel function and performs linear regression in that space. In this embodiment, the SVR model uses a radial basis function kernel. The kernel function parameter σ and the penalty coefficient C are optimized and selected through cross-validation.

[0049] The other steps in this embodiment are the same as in Embodiment 1, and will not be repeated here. Example

[0050] This embodiment provides a power load forecasting system based on adaptive time-series decomposition, which is used to implement the steps in the above method embodiment.

[0051] The system includes a data acquisition module, an adaptive decomposition module, a component prediction module, and a reconstruction module.

[0052] The data acquisition module is used to acquire historical power load data and preprocess it. Specifically, this module includes a data acquisition submodule, a missing value processing submodule, an outlier detection submodule, and a data normalization submodule. The data acquisition submodule connects to the power system data acquisition and monitoring control system via a data interface to acquire historical load data in real time or periodically read it. The missing value processing submodule fills in missing data using linear interpolation or nearest-neighbor filling. The outlier detection submodule identifies and corrects outliers based on the statistical distribution characteristics or rate of change thresholds of the load data. The data normalization submodule maps the load data to a unified data range using a max-min normalization method.

[0053] The adaptive decomposition module is used to construct an adaptive time-series decomposition model, which decomposes the preprocessed load data into several sub-series components. The number of decomposition levels in the adaptive time-series decomposition model is adaptively determined based on the characteristic parameters of the load data. This module includes a decomposition parameter adaptive determination submodule and a time-series decomposition execution submodule. The decomposition parameter adaptive determination submodule performs trial decomposition of the load data layer by layer according to the center frequency distribution characteristics and modal aliasing evaluation index of the load data, and determines the optimal number of decomposition levels. The time-series decomposition execution submodule performs the final decomposition of the load data according to the optimal number of decomposition levels and outputs each sub-series component.

[0054] The component prediction module performs stationarity tests on each subsequence component, matches a corresponding prediction model to each subsequence component based on the test results, and uses the matched prediction model to predict each subsequence component, thus obtaining the prediction result for each subsequence component. This module includes a stationarity test submodule, a prediction model matching submodule, and a component prediction execution submodule. The stationarity test submodule uses the unit root test method to perform stationarity tests on each subsequence component. The prediction model matching submodule matches a prediction model of the appropriate type to each component based on the stationarity test results. The component prediction execution submodule uses the matched model to predict each component and outputs the prediction result.

[0055] The reconstruction module integrates and reconstructs the prediction results of each subsequence component to obtain the power load prediction result. This module includes an initial reconstruction submodule and an error correction submodule. The initial reconstruction submodule linearly superimposes the prediction results of each subsequence component to obtain the initial reconstruction result. The error correction submodule corrects the initial reconstruction result based on the statistical characteristics of historical prediction errors or based on the predicted values ​​of the error sequence, and outputs the final prediction result. Example

[0056] This embodiment provides an electronic device, which includes a memory, a processor, and a computer program stored in the memory and executable on the processor.

[0057] The memory can be a storage medium such as random access memory, read-only memory, flash memory, or hard disk, used to store computer programs and related historical power load data, intermediate calculation results, and prediction results. The memory stores computer program instructions that implement the steps in the above method embodiments.

[0058] The processor can be a central processing unit, digital signal processor, application-specific integrated circuit or field-programmable gate array, etc., to execute computer programs in memory; when executing the program, the processor implements the various steps in the above method embodiments, including acquiring historical power load data and preprocessing it, constructing an adaptive time series decomposition model for decomposition, performing stationarity checks on each subsequence component and matching the prediction model for prediction, and integrating and reconstructing the prediction results of each subsequence component.

[0059] The electronic device also includes input / output interfaces for connecting external devices to input and output data; these interfaces include, but are not limited to, universal serial bus interfaces, network interfaces, and display interfaces. Through these interfaces, the electronic device can receive real-time load data from the power system and output the forecast results to the power dispatching system or display devices.

[0060] The working process of the electronic device is as follows: First, historical power load data is acquired through the input / output interface. The processor executes the program in the memory to preprocess the data. Then, an adaptive time series decomposition model is constructed to determine the optimal decomposition level and decompose the data into several subsequence components. After stationarity testing and prediction model matching of each component, predictions are made separately. Finally, the prediction results of each component are integrated and reconstructed, and the final prediction result is output through the input / output interface.

[0061] In this embodiment, the electronic device can be a desktop computer, server, workstation, or embedded industrial control computer, etc. In practical applications, the electronic device can be deployed in the power dispatch center or load forecasting center and connected to the power system data acquisition system through a network to realize automated load forecasting function.

[0062] The power load forecasting method, system, and electronic device based on adaptive time-series decomposition provided in the above embodiments of the present invention achieve refined decomposition and targeted forecasting of power load data by adaptively determining the number of time-series decomposition layers and matching corresponding forecasting models to each component, thereby effectively improving forecasting accuracy. Those skilled in the art will understand that all or part of the steps of the above method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium, and when executed, it performs the steps of the above method embodiments.

[0063] In summary, this invention can adaptively process load data with different characteristics, significantly improving the accuracy and stability of power load forecasting, and is applicable to power load forecasting scenarios at different time scales.

[0064] It is understood that the present invention has been described through some embodiments, and those skilled in the art will recognize that various changes or equivalent substitutions can be made to these features and embodiments without departing from the spirit and scope of the invention. Furthermore, under the teachings of the present invention, these features and embodiments can be modified to adapt to specific situations and materials without departing from the spirit and scope of the invention. Therefore, the present invention is not limited to the specific embodiments disclosed herein, and all embodiments falling within the scope of the claims of this application are within the protection scope of the present invention.

Claims

1. A power load forecasting method based on adaptive time-series decomposition, characterized in that, Includes the following steps: S1. Obtain historical power load data and preprocess the historical power load data; S2. Construct an adaptive time series decomposition model to decompose the preprocessed load data into several subsequence components. The number of decomposition layers of the adaptive time series decomposition model is adaptively determined based on the characteristic parameters of the load data. S3. Perform stationarity tests on each subsequence component, match the corresponding prediction model to each subsequence component based on the test results, and use the matched prediction model to predict each subsequence component to obtain the prediction results of each subsequence component. S4. Integrate and reconstruct the prediction results of each subsequence component to obtain the power load prediction result.

2. The power load forecasting method based on adaptive time-series decomposition according to claim 1, characterized in that, The preprocessing in step S1 includes: Historical power load data is processed for missing values, outlier detection and correction, and data normalization. The missing value processing adopts linear interpolation or neighbor-value filling methods, and the outlier detection is based on the statistical distribution characteristics or rate of change threshold of the load data.

3. The power load forecasting method based on adaptive time-series decomposition according to claim 1, characterized in that... The adaptive time-series decomposition model is constructed based on the variational mode decomposition algorithm. The number of decomposition layers is adaptively determined according to the center frequency distribution characteristics of the load data and the evaluation index of mode aliasing, including: The load data is decomposed layer by layer according to the preset decomposition layer range. The ratio of the modal center frequency distribution under each decomposition layer is calculated. When the ratio of the center frequencies of adjacent modes is less than the preset frequency threshold, the decomposition layer is determined as the optimal decomposition layer.

4. The power load forecasting method based on adaptive time-series decomposition according to claim 1, characterized in that, The adaptive time series decomposition model is constructed based on the ensemble empirical mode decomposition algorithm. The number of decomposition layers is adaptively determined according to the extreme point distribution density and energy loss rate of the load data. The extreme point distribution density is calculated according to the formula =, where M is the total number of local extreme points in the load time series and N is the sequence length. The energy loss rate is calculated according to the formula =, where is the original load signal energy, is the energy of the Kth modal component, and K is the decomposition layer number; K is increased layer by layer and the corresponding value is calculated. When the decrease rate of K is less than the preset threshold for the first time, the current K is determined as the optimal decomposition layer number.

5. The power load forecasting method based on adaptive time-series decomposition according to claim 1, characterized in that, In step S3, stationarity tests are performed on each subsequence component, and corresponding prediction models are matched to each subsequence component based on the test results, including: The unit root test method was used to test the stationarity of each subsequence component. Subsequence components that pass the stationarity test are identified as stationary components, and subsequence components that fail the stationarity test are identified as non-stationary components. The stationary component is matched with a linear prediction model, and the non-stationary component is matched with a nonlinear prediction model.

6. The power load forecasting method based on adaptive time-series decomposition according to claim 5, characterized in that, The linear prediction model is an autoregressive integral moving average model or an autoregressive model, and the nonlinear prediction model is a long short-term memory network model, a gated recurrent unit model, or a support vector regression model.

7. The power load forecasting method based on adaptive time-series decomposition according to claim 1, characterized in that... Step S4 involves integrating and reconstructing the prediction results of each sub-sequence component, including: The prediction results of each subsequence component are linearly superimposed to obtain the initial reconstruction result; The initial reconstruction result is subjected to error correction processing, which is based on the statistical characteristics of historical prediction errors or the predicted values ​​of the error sequence.

8. The power load forecasting method based on adaptive time-series decomposition according to claim 1, characterized in that... The characteristic parameters include at least one of the following: the fluctuation frequency of the load data, the range of amplitude variation, and the trend strength.

9. A power load forecasting system based on adaptive time-series decomposition, characterized in that, include: The data acquisition module is used to acquire historical power load data and preprocess the historical power load data; The adaptive decomposition module is used to construct an adaptive time-series decomposition model, which decomposes the preprocessed load data into several sub-series components. The number of decomposition layers of the adaptive time-series decomposition model is adaptively determined based on the characteristic parameters of the load data. The component prediction module is used to perform stationarity tests on each subsequence component, match the corresponding prediction model for each subsequence component based on the test results, and use the matched prediction model to predict each subsequence component to obtain the prediction result of each subsequence component. The reconstruction module is used to integrate and reconstruct the prediction results of each subsequence component to obtain the power load prediction result.

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

Images