A short-term power load prediction method based on modal fourier decomposition and improved Autoformer

Abstract

The present application relates to a kind of short-term power load prediction method based on modal fourier decomposition and improved Autoformer, belong to electric power technical field.The method of the present application includes three parts of influence factor correlation analysis module, modal fourier joint decomposition module and integrated module of improved Autoformer.The short-term power load prediction method proposed in the present application has higher prediction accuracy and stronger robustness under the condition that load presents complex nonlinear characteristics, noise interference and obvious seasonal fluctuation.Load decomposition test, robustness test, one week test and one day test are carried out, and the experimental results show that modal fourier joint decomposition module significantly reduces the complexity of load component, the method proposed in the present application has better robustness than traditional model, has higher stability in one week and one day test, and has the highest prediction accuracy.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of electric power and relates to a short-term electric power load prediction method based on modal Fourier decomposition and an improved Autoformer. BACKGROUND

[0002] Accurate short-term electric power load prediction is crucial for the stable and economic operation of power grids. However, electric power load is influenced by user behavior, weather, holidays, and other external factors, among which weather factors have a greater impact. In recent years, the frequency of extreme weather events has increased, and their intensity has intensified. People's demand for cooling and heating has increased dramatically, leading to a surge in electric power load demand, resulting in strong fluctuations in electric power load and a large amount of noise, which poses a great challenge to short-term electric power load prediction.

[0003] Existing electric power load prediction methods mainly include traditional statistical methods, machine learning methods, neural network methods, and combined model methods. Traditional statistical methods, such as autoregressive models (AR), autoregressive integrated moving average models (ARIMA), and exponential smoothing methods, have simple structures and fast training speeds, but they are difficult to capture nonlinear trends in electric power load and are sensitive to noise. Machine learning methods include random forests, gradient boosting trees, and support vector machines. Compared with traditional statistical methods, machine learning methods can effectively handle nonlinear features in data and exhibit stronger generalization ability, but they are usually suitable for scenarios with low complexity of load prediction data. Neural network methods, such as long short-term memory networks (LSTM), temporal convolutional networks (TCN), Transformers, and Autoformers, have their own advantages and disadvantages. LSTM handles time series through recursion, but it has the problem of gradient explosion. TCN has good ability to capture time series features, but it lacks the ability to capture long-term features. Transformers have strong nonlinear mapping ability, but they are inefficient in handling long-term time series. Autoformers have stronger feature extraction ability than Transformers, but they lack convolution, which affects their ability to capture short-term features.

[0004] Many studies propose the method of combining models, such as: a combination prediction model combining K-means and extreme learning machine, which clusters user behavior characteristics by K-means to reduce the dimension and complexity of feature data, and uses extreme learning machine for prediction; a VMD-VAE-LSTM hybrid load prediction model, which decomposes the data into multiple subsequences, extracts features from the decomposed sequences using VAE, and finally uses LSTM for prediction; a model combining CEEMDAN (Complete Ensemble Empirical Mode Decomposition with Adaptive Noise) and sample entropy and Transformer, which decomposes the original load sequence and combines Transformer for prediction. The above methods do not conduct in-depth analysis on the correlation between load components, and do not fully consider the influencing factors, thereby causing the dimension of the load component to be too high, increasing the calculation difficulty. Due to the scale limitation of the model for feature extraction, the extraction of local features is not comprehensive, which further affects the robustness and prediction accuracy of the prediction model. SUMMARY

[0005] Therefore, the present application provides a short-term power load prediction method based on mode Fourier decomposition and improved Autoformer (Mode Fourier-Improved Autoformer Joint Prediction Method, MF-IAFP), which solves the problems of insufficient robustness and low prediction accuracy of existing short-term power load prediction methods.

[0006] To achieve the above-mentioned purpose, the present application provides the following technical solutions:

[0007] A short-term power load prediction method based on mode Fourier decomposition and improved Autoformer, the method comprising the following steps:

[0008] S1: Use the Spearman correlation coefficient method to analyze the correlation between the load and other influencing factors, and screen out the key influencing factors;

[0009] S2: Combine CEEMDAN (Complete Ensemble Empirical Mode Decomposition with Adaptive Noise) and FFT (Fast Fourier Transform) to propose a mode Fourier joint analysis method (Mode Fourier Joint Analysis, MFJA), which realizes the decomposition of the original power load;

[0010] S3: Establishing a load component dataset of key influencing factors and load components;

[0011] S4: Introducing a temporal convolutional network (TCN) to improve the autocorrelation module in Autoformer;

[0012] S5: Establishing an integrated IAutoformer model for each load component dataset and performing prediction;

[0013] S6: Fusing the prediction values of each model to output the final prediction result.

[0014] Further, the S1 comprises the following steps:

[0015] 1) Collecting power load data through a power load monitoring platform;

[0016] 2) Obtaining meteorological factor data of temperature, wind speed, rainfall, and humidity through a meteorological platform;

[0017] 3) Performing missing value processing, outlier processing, data standardization, and normalization on the collected data;

[0018] 4) Using independent coding to encode weekdays and hours;

[0019] 5) Using the Spearman correlation coefficient method to perform correlation analysis on the load data and each influencing factor, and selecting the influencing factors related to the load.

[0020] Further, the modal Fourier joint analysis method comprises the following steps:

[0021] 1) Inputting the original power load sequence into a completely adaptive noise ensemble empirical mode decomposition model to obtain load components of different frequencies;

[0022] 2) Using fast Fourier transform to obtain the main frequency of each component;

[0023] 3) Setting a frequency difference threshold Δf;

[0024] 4) Calculating the frequency difference value between each pair of components, and if the difference value is less than Δf, then optimizing and fusing them;

[0025] 5) Finally outputting the new load component.

[0026] Further, the S3 comprises the following specific steps:

[0027] According to the selected key influencing factors and the output load components, load component datasets of each load component are established.

[0028] Further, the S4 comprises the following steps:

[0029] 1) After linear transformation of the query matrix Q and the key matrix K, the frequency domain representation is obtained by FFT, and the frequency domain feature Z is obtained by splicing f ;

[0030] 2) Linear transformation is performed on the value matrix V to obtain V1;

[0031] 3) The frequency domain feature Z f is transformed back to the time domain by inverse fast Fourier transform (IFFT) to obtain the time domain feature Z t ;

[0032] 4) Delay aggregation is performed on Z t , V1 to obtain the time domain feature Z g after delay aggregation, and Z c is the spliced result;

[0033] 5) The spliced feature is subjected to multi-scale feature extraction through the TCN structure, and is output through the linear layer.

[0034] Further, the integrated IAutoformer model is established, comprising the following steps:

[0035] 1) According to the established load component data set, the IAutoformer model is established respectively;

[0036] 2) The load component data set is divided into a training set and a test set;

[0037] 3) The training of each model is performed and the parameters are adjusted.

[0038] Further, the prediction values of each model are fused to output the final prediction result, comprising the following steps:

[0039] According to the established integrated IAutoformer model, the test set is input into each model, and the final output prediction result is obtained by fusing the prediction output values of each model.

[0040] The beneficial effects of the present application are as follows:

[0041] 1. The influence factor correlation analysis module is used to analyze the correlation between external factors such as temperature, humidity, wind speed, time and short-term power load, and to screen out key influencing factors, thereby significantly reducing the calculation complexity of model prediction.

[0042] 2. The modal components processed by the modal Fourier joint decomposition module have significantly reduced modal complexity compared to the modal components obtained by CEEMDAN decomposition, and effectively reduced the fluctuation of low- and medium-frequency components.

[0043] 3. Compared with other short-term power load forecasting models, the method proposed in this invention exhibits the highest prediction accuracy across different datasets. Furthermore, the method proposed in this invention outperforms other traditional models in terms of anti-interference capability and stability, demonstrating the best robustness.

[0044] Other advantages, objectives, and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination, or may be learned from practice of the invention. The objectives and other advantages of the invention can be realized and obtained through the following description. Attached Figure Description

[0045] To make the objectives, technical solutions, and advantages of the present invention clearer, the preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings, wherein:

[0046] Figure 1 The structure and data flow diagram of TCN provided in the embodiments of the present invention;

[0047] Figure 2 This is an improved autocorrelation module structure provided in the embodiments of the present invention;

[0048] Figure 3 This is a block diagram of the overall architecture of the MF-IAFP model provided in an embodiment of the present invention;

[0049] Figure 4 A heatmap of correlation coefficients of influencing factors provided in an embodiment of the present invention;

[0050] Figure 5 The MFJA test result diagram provided in the embodiment of the present invention;

[0051] Figure 6 This is a chart showing the test results for each dataset provided in this embodiment of the invention;

[0052] Figure 7 Box plots of one-week test error for each dataset provided in embodiments of the present invention;

[0053] Figure 8 A graph showing the test results for each dataset provided in this embodiment of the invention;

[0054] Figure 9 Box plots of one-day test error for each dataset provided in embodiments of the present invention. Detailed Implementation

[0055] Other advantages and benefits of the present application will become apparent to those skilled in the art upon reading and understanding the following detailed description of the application. The present application may be carried out in other different embodiments and implemented in other different ways without departing from the spirit of the present application. Each of the details of the present application may be modified or changed based on different views and applications without departing from the spirit of the present application. It should be noted that the drawings provided in the following embodiments only schematically illustrate the basic concept of the present application, and the following embodiments and features in the embodiments can be combined with each other without conflict.

[0056] The drawings are only used for illustrative description, and the representation is only a schematic diagram, not a physical diagram, and should not be understood as a limitation of the present application. In order to better illustrate the embodiments of the present application, some components in the drawings may be omitted, enlarged or reduced, and do not represent the actual size of the product. It is understandable for those skilled in the art that some well-known structures and their descriptions in the drawings may be omitted.

[0057] The same or similar reference numerals in the drawings of the embodiments of the present application correspond to the same or similar components; in the description of the present application, it should be understood that if the terms "upper", "lower", "left", "right", "front", "back" and the like indicate the orientation or positional relationship shown in the drawings, they are only for the convenience of describing the present application and simplifying the description, and do not indicate or imply that the devices or elements referred to must have a particular orientation, be constructed and operated in a particular orientation, therefore the terms describing the positional relationship in the drawings are only used for illustrative description, and should not be understood as a limitation of the present application. For those skilled in the art, the specific meaning of the above terms can be understood according to the specific circumstances.

[0058] The present application provides a short-term power load prediction method based on modal Fourier decomposition and improved Autoformer. The method mainly includes an influence factor correlation analysis module, a modal Fourier joint decomposition module and an improved Autoformer integration module to complete short-term power load prediction.

[0059] The present embodiment first makes a whole description of the basic principles and execution process of the above three parts, and then the specific steps of each part are as follows:

[0060] Step one: correlation analysis of influence factors

[0061] Through the power real-time monitoring platform and the meteorological collection platform, power load data, temperature, rainfall, humidity, wind speed and other meteorological factors are collected, and the collected data are preprocessed, including abnormal data detection, missing value filling, data standardization and normalization processing, so as to improve the accuracy of correlation calculation between each data. The Spearman correlation coefficient method is used to analyze the correlation between load and each external influencing factor, and the high correlation influencing factors are selected as the input features of the subsequent model.

[0062] Step two: modal Fourier combined analysis method decomposes load

[0063] The load sequence is decomposed by CEEMDAN to obtain a plurality of stationary subsequences (IMF1, IMF2, …, IMFn) with different frequency characteristics. Next, FFT is used to extract the dominant frequency of each load component, and according to the set frequency difference threshold, the coupling processing of the frequency similar modes is performed, and finally the processed load component is output; the execution process is as follows:

[0064] 1) Input the original power load sequence x(t), and use the CEEMDAN algorithm to decompose x(t) to obtain a series of IMF components IMFi (where i = 1, 2, …, n).

[0065] 2) Apply FFT to each IMF component to obtain the frequency spectrum F i (f) represents, and the dominant frequency f i of each IMF i component is determined.

[0066] 3) Set the frequency difference threshold Δf, if the dominant frequency difference of two IMF components is less than the set threshold, it is considered that they can be fused.

[0067] 4) Fusion rule: for adjacent IMF i and IMF i+1 , if |f i -f i+1 | < Δf, then they are fused, and the fused new IMF component is denoted as FIMF j (where j = 1, 2, …, m), and the calculation formula is:

[0068] FIMF j = IMF i + IMF i+1

[0069] 5) The fused component FIMF j (where j = 1, 2, …, m) is taken as the final result of the original power load decomposition and component fusion, and is input to the data set of fused influencing factors.

[0070] Step three: improving the integrated module of Autoformer

[0071] The influence factors and modal Fourier joint decomposition module output load components are constructed based on the influence factor correlation analysis module screening. For each load component data set, an independent improved Autoformer prediction model is established, and each load component is predicted. Finally, the prediction results of each model are reconstructed and fused to obtain the overall load prediction result. The execution process is as follows:

[0072] 1) Constructing TCN network

[0073] The TCN structure is shown in Figure 1 , which consists of two serially connected sub-modules. Each sub-module contains an expanded causal convolution layer, a Dropout layer, a ReLU activation function, and a weight normalization (WeightNorm) component. The expansion factor of the first layer of expanded causal convolution is set to 1, which is mainly used to extract short-term dependence features; the expansion factor of the second layer is set to 2, which is used to capture longer-term dependence features.

[0074] 2) Improved autocorrelation module

[0075] The improved autocorrelation module structure is shown in Figure 2 , which performs time convolution operation on the output of the original autocorrelation module to extract the complex relationship between local load and load, load and influence factors. The data processing steps of the improved autocorrelation module are as follows:

[0076] After linear transformation of the query matrix Q and the key matrix K, the frequency domain representation Q1 of the query matrix and the frequency domain representation K1 of the key matrix are obtained by FFT respectively;

[0077]

[0078] The frequency domain feature Z f is represented as:

[0079]

[0080] In the formula, represents frequency domain multiplication.

[0081] The Z f is subjected to inverse fast Fourier transform (IFFT) to obtain the time domain feature Z t :

[0082] Z t = IFFT(Z f )

[0083] Linearly transform the value matrix V to obtain V1, and perform time delay aggregation on Z t and V1 to obtain the aggregated feature Z g ;

[0084] V1 = Linear(V)

[0085] Z g = TimeDelayAgg(Z t , V1)

[0086] Concatenate Z g after splicing to obtain the spliced feature Z c .

[0087] Z c = Concat(Z g )

[0088] Z c After multi-scale feature extraction by TCN, output through a linear layer, and obtain the improved autocorrelation module output result Z out .

[0089] Z out = TCN(Z c )

[0090] 3) Establish the MF-IAFP model

[0091] Figure 3 is the overall architecture diagram of the MF-IAFP model, including influence factor analysis, modal Fourier joint analysis method, and establishment of an integrated short-term power load prediction model. Among them, the integrated short-term power load prediction model constructs a load component dataset, establishes an independent improved Autoformer prediction model for each load component dataset, and performs load component prediction. Finally, the prediction results of each model are reconstructed and fused to obtain the overall load prediction result.

[0092] Based on the above three parts, the short-term power load prediction method based on modal Fourier decomposition and improved Autoformer provided in the embodiment includes the following steps:

[0093] Step 1: Perform correlation analysis of influence factors

[0094] Step 1.1: Collect power load data, temperature, rainfall, humidity, wind speed and other meteorological factors through the power real-time monitoring platform and the meteorological collection platform.

[0095] Step 1.2: Preprocess the collected data, including abnormal data detection, missing value filling, data standardization and normalization processing, so as to improve the accuracy of correlation calculation between data.

[0096] Step 1.3: Analyze the correlation between load and external factors using the Spearman correlation coefficient method, and screen high-correlation factors as input features for subsequent models.

[0097] This example collects three different geographical data sets as research objects, of which data set 1 is the actual load data set collected in a certain place in southwest China, with a time interval of 15 minutes; data set 2 is a public data set in a certain place in central China, with a time interval of 15 minutes; data set 3 is a public data set in a certain place in Australia, with a time interval of 0.5 hours; Taking data set 1 as an example, the influence factor screening, modal Fourier joint analysis method testing and robustness testing are carried out. In order to verify the stability and prediction accuracy of MF-IAFP and other models on different time scales and different data sets, this paper carries out one-week prediction test and one-day prediction test on the above data sets. The training rounds (epochs) of each model in the experiment are 100, the optimizer is selected as Adam, the learning rate is set to 0.0005, and the dropout rate is 0.3, which ensures the comparison of the performance of each model under the same experimental conditions.

[0098] The correlation between each factor and load is analyzed by using the Spearman correlation coefficient. The absolute value of the correlation coefficient is between 0.8 and 1 for extremely strong correlation, between 0.6 and 0.8 for strong correlation, between 0.4 and 0.6 for moderate intensity correlation, between 0.2 and 0.4 for weak correlation, and between 0 and 0.2 for no correlation. As shown in Table 1, the correlation coefficients of Temperature, Rainfall, Month, Hour and Load are 0.67, 0.36, 0.25 and 0.42, respectively, so they are used as input features for establishing short-term power load prediction models. Figure 4

[0099] Step 2: Decompose the power load using modal Fourier joint analysis method

[0100] Step 2.1: Input the preprocessed power load data into CEEMDAN for decomposition to obtain power load components of different frequencies;

[0101] Step 2.2: Extract the main frequency in each load component using FFT;

[0102] Step 2.3: Set Δf to select and fuse components with small frequency differences for optimization, reduce the complexity of modal components, and output the final decomposition results.

[0103] The load time series of data set 1 for 15 consecutive days are selected for analysis, Figure 5 ​The results of the MFJA test are shown. The method decomposes the original load time series data into 8 intrinsic mode components (IMFs) using CEEMDAN. IMF1 and IMF2 are high-frequency components that mainly capture the noise and short-term changes in the load sequence, reflecting the strong random fluctuation characteristics in the load; IMF3 to IMF5 are medium-frequency components that reveal the daily fluctuation rules in the load; IMF6 to IMF8 are low-frequency components that reflect the long-term trend and seasonal changes in the load sequence. Second, the main frequency of each modal component is analyzed by FFT as shown in Table 1, and the coupling components in the load component are processed by setting Δf, wherein IMF4 and IMF5 are merged into FIMF4, and IMF6 and IMF7 are merged into FIMF5. After the MFJA processing of the modal components, compared with the modal components obtained by CEEMDAN decomposition, the modal complexity is significantly reduced, and the volatility of the medium and low frequency components is effectively reduced.

[0104] Table 1

[0105] component IMF1 IMF2 IMF3 IMF4 IMF5 IMF6 IMF7 IMF8 frequency 0.26041 0.125 0.03125 0.01041 0.01041 0.00138 0.00138 0.00001

[0106] Step 3: Establishing the MF-IAFP model

[0107] Step 3.1: Constructing the load component dataset based on the influence factors selected by the influence factor correlation analysis module and the load components output by the modal Fourier combined decomposition module;

[0108] Step 3.2: Establishing an IAutoformer prediction model for each load component dataset;

[0109] Step 3.3: Performing short-term power load prediction;

[0110] Step 3.4: Fusing the prediction values of each model output value to obtain the final prediction result.

[0111] One-week test: The MF-IAFP model proposed in the present application is compared and analyzed with LSTM, CNN, Transformer, Autoformer, IAutoformer and MF-Autoformer on different datasets. Figure 6The one-week prediction results of each model on different datasets are shown. On dataset 1, the prediction effect of CNN and Transformer is poor, and the prediction curves of other models are close to the true value. By magnifying a wave trough, it can be found that the MF-IAFP model performs well in capturing subtle fluctuations in load, and the prediction result is closest to the true value. On dataset 2, CNN, Autoformer, Transformer and IAutoformer have obvious deviation in predicting wave peaks and wave troughs. In contrast, LSTM, MF-Autoformer and MF-IAFP have better prediction results. In particular, when magnifying a wave peak, the prediction curve of MF-IAFP is highly consistent with the true value, showing strong prediction ability. On dataset 3, Autoformer's prediction result deviates significantly from the true value at a wave trough, while other models have relatively good prediction results. When magnifying a double-peak region, the prediction value of MF-IAFP is closest to the true value, further verifying its superiority in complex fluctuation patterns. The MF-IAFP model exhibits stronger prediction ability on all datasets, especially in capturing subtle fluctuations in load time series.

[0112] Figure 7 The one-week test error box plots of each dataset are shown. From the figure, it can be seen that the box height of the MF-IAFP model in each dataset is the lowest, indicating that the distribution of its prediction error is the most concentrated.

[0113] Table 2 shows the one-week test index results of each model on different datasets. In dataset 1 test, the MAPE of MF-IAFP is 0.552, the RMSE is 57.902, the MAE is 45.267, the R 2 is 0.9978, which is 4%, 5.37%, 3.23%, and 0.03% higher than the other optimal models, respectively. In dataset 2 test, the MAPE of MF-IAFP is 0.664, the RMSE is 577.587, the MAE is 434.113, the R 2 is 0.9971, which is 41.91%, 39.03%, 42.87%, and 0.49% higher than the other optimal models, respectively. In dataset 3, the MAPE of MF-IAFP is 0.424, the RMSE is 1919.863, the MAE is 1543.969, the R 2 is 0.9837, which is 11.48%, 10.85%, 11.65%, and 0.43% higher than the other optimal models, respectively.

[0114] Table 2

[0115]

[0116] One-day test: The present application conducts one-day test comparative analysis on each model on different data sets, and the results are shown in Figure 8 Figure 2. In data set 1, the prediction result of MF-IAFP is the most accurate in the 0-4 hour interval, and other models have large deviations; in the 8-12 hour interval, the prediction curve of MF-IAFP closely follows the fluctuation of the true value, while the prediction curves of other models are relatively smooth and have large deviations. In data set 2, the prediction curves of Transformer, Autoformer and MF-Autoformer in some time periods are seriously deviated from the true value, and the prediction results of LSTM, CNN and MF-IAFP are relatively good, among which the prediction result of MF-IAFP is the best. In data set 3, the prediction results of each model are generally good. After amplifying the analysis of the 12-14 hour interval, it is found that the prediction result of MF-IAFP is the most accurate.

[0117] Figure 9 The one-day test error box plot of each data set is shown. Consistent with the one-week test results, the box height of the MF-IAFP model is the smallest, indicating that its error value distribution is the most concentrated.

[0118] Table 3 shows the one-day test results of each model on three data sets. In data set 1 test, the MAPE of MF-IAFP is 0.439, the RMSE is 46.360, the MAE is 36.501, the R 2 is 0.9981, which is improved by 33.58%, 34.12%, 34.73% and 0.25% compared with other optimal models respectively. In data set 2, the MAPE of MF-IAFP is 0.807, the RMSE is 559.041, the MAE is 491.804, the R 2 is 0.9959, which is improved by 13.87%, 35.93%, 17.47% and 0.58% compared with other optimal models respectively. From data set 3, the MAPE of MF-IAFP is 0.296, the RMSE is 1406.81, the MAE is 1088.799, and the R 2 reaches 0.9927. Compared with other optimal models, MF-IAFP improves by 36.09%, 34.22%, 35.49% and 0.9% respectively in each index, fully showing its excellent prediction accuracy.

[0119] Table 3

[0120]

[0121] Robustness test: Based on dataset 1, the robustness of each model was tested, and three cases were considered: common random noise (Gaussian noise), extreme interference (uniform noise), and ideal conditions (no noise). By comparing the three cases, the robustness and adaptability of each model can be comprehensively evaluated. The test results are shown in Table 4. Under the condition of no noise, the MAPE of MF-IAFP is 0.296, the RMSE is 1406.818, the MAE is 1088.799, and the R 2 value is 0.9927. Under Gaussian noise interference, the MAPE of MF-IAFP is 0.409, the RMSE is 1828.211, the MAE is 1495.68, and the R 2 value is 0.9877. Under uniform noise interference, the MAPE of MF-IAFP is 0.320, the RMSE is 1505.592, the MAE is 1178.789, and the R 2 value is 0.9917. Comparing the differences in MAPE, RMSE, and MAE of each model under Gaussian noise, uniform noise, and no noise conditions, the results are shown in Table 5. MF-IAFP shows the smallest value in each difference: the difference in MAPE is 0.113, 0.024, the difference in RMSE is 421.393, 98.774, and the difference in MAE is 406.881, 89.99. These results show that the MF-IAFP model has strong robustness and can maintain excellent prediction performance in complex noise environments.

[0122] Table 4

[0123]

[0124] Table 5

[0125]

[0126]

[0127] Finally, it should be noted that the above examples are only used to illustrate the technical solutions of the present application and are not limiting. Although the present application has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present application can be modified or replaced by equivalents without departing from the spirit and scope of the present application, and all such modifications or replacements should be included in the scope of the claims of the present application.

Claims

1. A short-term power load prediction method based on modal Fourier decomposition and improved Autoformer, characterized by: The method comprises the following steps: S1: correlation analysis of load and other influencing factors is performed by using the Spearman correlation coefficient method, and key influencing factors are screened out; S2: a mode Fourier joint analysis method (MFJA) is proposed by combining complete ensemble empirical mode decomposition with autoencoding (CEEMDAN) and fast Fourier transform (FFT), so as to realize decomposition of original power load; S3: a load component dataset of key influencing factors and load components is established; S4: a time convolution network (TCN) is introduced to improve the autocorrelation module in the Autoformer; S4 comprises the following steps: 1) Linearly transform the query matrix Q and the key matrix K to get the frequency domain representation by FFT, and splice to get the frequency domain feature Z f ; 2) the value matrix V is linearly transformed to obtain V 1; 3) Frequency domain features Z f The time-domain features are obtained by using Inverse Fourier Transform (IFFT) to return to the time domain. Z t ; 4) on Z t , V 1 delay aggregation, to get the time domain features after delay aggregation Z g , Z c is the result of splicing; 5) the spliced features are subjected to multi-scale feature extraction through the TCN structure, and are output through a linear layer; S5: an integrated IAutoformer model is established for each load component dataset, and prediction is performed; S6: the prediction values of the models are fused, and the final prediction result is output.

2. The short-term power load forecasting method based on modal Fourier decomposition and improved Autoformer according to claim 1, characterized in that: S1 comprises the following steps: 1) power load data is collected through a power load monitoring platform; 2) meteorological factor data such as temperature, wind speed, rainfall and humidity are obtained through a meteorological platform; 3) missing value processing, abnormal value processing, data standardization and normalization are performed on the collected data; 4) independent coding is used to code working days and hours; 5) correlation analysis of load data and each influencing factor is performed by using the Spearman correlation coefficient method, and the influencing factors related to the load are screened out. 3.The short-term power load forecasting method based on modal Fourier decomposition and improved Autoformer according to claim 1, wherein: The mode Fourier joint analysis method comprises the following steps: 1) the original power load sequence is input into a complete ensemble empirical mode decomposition model with autoencoding (CEEMDAN), and load components of different frequencies are obtained; 2) the main frequency of each component is obtained by using fast Fourier transform (FFT); 3) setting a frequency difference threshold value Δ f ; 4) Calculate the frequency difference value between each pair of components, if the difference value is less than Δ f then fuse and optimize them; 5) finally, the new load component is output.

4. The short-term power load forecasting method based on modal Fourier decomposition and improved Autoformer according to claim 2, characterized in that: The specific steps of S3 are as follows: According to the screened key influencing factors and the output load components, a load component dataset of each load component is established.

5. The short-term power load forecasting method based on modal Fourier decomposition and improved Autoformer according to claim 4, characterized in that: The establishment of the integrated IAutoformer model comprises the following steps: 1) an IAutoformer model is established according to the established load component dataset; 2) the load component dataset is divided into a training set and a test set; 3) each model is trained and the parameters are adjusted.

6. The short-term power load forecasting method based on modal Fourier decomposition and improved Autoformer according to claim 5, characterized in that: The prediction values of the models are fused, and the final prediction result is output, comprising the following steps: According to the established integrated IAutoformer model, the divided test set is input into each model, and the final output prediction result is obtained by fusing the prediction output values of each model.

Images