Power load forecasting system and method based on EFformer model

Denoising and feature extraction of time series loads through the encoder and decoder module of the EFformer model, solving the problem of noise and periodic feature extraction in time series load prediction in power system, and achieving accurate prediction of volatility loads , improves the operating stability and scheduling optimization of the power system.

CN118659367BActive Publication Date: 2024-11-26STATE GRID TIANJIN ELECTRIC POWER CO CHENGXI POWER SUPPLY BRANCH +1
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411124588.1
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-08-16
Publication Date
2024-11-26
Estimated Expiration
2044-08-16

AI Technical Summary

Technical Problem

The prior art has problems in data noise, difficulty in extracting periodic features and lag in volatility load prediction in power systems and smart grids, resulting in insufficient prediction accuracy and reliability.

Method used

Using the power load prediction system based on the EFformer model, denoising and feature extraction of time series loads is performed through encoder and decoder modules, and using wavelet transform and multi-subband decomposition technology to remove noise, extract trends and periodic features, and use Transformer to Structural processing timing characteristics.

Benefits of technology

It realizes fast and accurate prediction of power load, improves the accuracy and reliability of prediction, can effectively capture time domain information and frequency domain characteristics, and is suitable for volatility and non-obvious periodic time series loads.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118659367B_ABST
    Figure CN118659367B_ABST
Patent Text Reader

Abstract

The present invention discloses an electric load forecasting system and method based on an EFformer model, belonging to the technical field of smart grids, comprising: a human-computer interaction module, used for receiving or storing a time series load of electric load data, and also used for outputting an electric load forecasting result; an EFformer model, which processes the time series load to obtain an electric load forecasting result; the EFformer model comprises an encoder and a decoder; a decoder attention module receives an output result of a multi-subband decomposition module of the encoder; the present invention utilizes the EFformer model to perform intelligent analysis on the time series load of the electric load, and then obtains an electric load forecasting result.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the technical field of smart grids, and in particular relates to a power load forecasting system and method based on an EFformer model. Background Art

[0002] As we all know, the energy system plays a vital role in my country's economic and social development. Nowadays, there are many types of energy, including coal, hydropower, oil, natural gas, nuclear energy, biomass energy, wind energy, solar energy, electric energy, etc. Among these energy sources, electric energy, as a major secondary energy source, is closely connected with other primary energy sources. It supplies electricity to various users through large-scale power systems, making energy utilization more convenient. After years of development, the power industry has become the central link of the modern energy system.

[0003] However, with the rapid development of industrial parks, the problem of energy waste has become increasingly serious. Especially in various industrial parks, energy consumption is unplanned, resulting in endless waste of electric energy. In order to use electric energy more efficiently, smart grids came into being.

[0004] Smart Grid and Artificial Intelligence Technology

[0005] In the past decade, researchers have conducted extensive research on smart grids (SG). The core feature of smart grids is the two-way flow between information and electricity. In this context, artificial intelligence (AI) technology is increasingly being used in smart grids, such as artificial neural networks (ANN), machine learning (ML), and deep learning (DL).

[0006] Among them, deep learning, as an emerging artificial intelligence technology, has been increasingly studied in the field of smart grids. However, the application of deep learning in smart grids is still immature and there is still a lot of research space and application potential.

[0007] Disadvantages of existing technology

[0008] In the current power system and smart grid field, time series load forecasting is a critical task. However, this task faces many challenges, including the following three aspects:

[0009] The first is the data noise problem. In smart grid SG and power systems, time series loads are often affected by a variety of factors, such as measurement errors, equipment failures, and abnormal events. These factors will introduce a lot of noise into the time series load data, causing the prediction results to lose accuracy and become unstable. Although the traditional moving average algorithm can reduce noise to a certain extent, it is less sensitive to the actual changes in the data, and the predicted values ​​often remain at the past level, unable to predict the high and low fluctuations of future loads.

[0010] The second is the problem of periodic feature extraction. Traditional deep learning DL models usually use Fourier transform to extract periodicity and trend features when predicting time series loads. However, the effectiveness of this method is based on the premise that the time series load is stable. In addition, Fourier transform is only a frequency domain transform, which can only reflect the frequency domain characteristics of the time series load, but cannot capture the time domain information. Therefore, this method has limitations when predicting time series loads with volatility and unclear periodicity.

[0011] The third is the problem of the long short-term memory recurrent neural network (LSTM) algorithm. Although the LSTM algorithm can maintain high accuracy when predicting simple linear and stable periodic signals (such as sine waves), its prediction results often lag significantly when facing time series loads with volatility and unclear periodicity. In addition, when the time series length is too long, LSTM may encounter problems such as long training time, slow parameter update, and even gradient disappearance during the prediction process.

[0012] In response to these problems, researchers are constantly exploring new methods and technologies to improve the accuracy and reliability of time series load forecasting. For example, by introducing more advanced algorithms and models, such as deep reinforcement learning and variational autoencoders, it is possible to better handle noisy data and capture the temporal information of time series loads. In addition, feature engineering and data preprocessing techniques can be used to reduce noise and enhance raw data to improve forecasting accuracy.

[0013] In short, time series load forecasting is of great significance in the field of power systems and smart grids. In order to solve the problems existing in the existing technologies, researchers need to continuously explore innovative methods to improve the accuracy and reliability of forecasting and provide strong support for the stable operation and optimal dispatch of power systems. Therefore, it is of great significance to design and develop a power load forecasting system and method based on the EFformer model. Summary of the invention

[0014] In order to solve the technical problems existing in the known technology, the present invention provides a power load forecasting system and method based on the EFformer model, which uses the EFformer model to perform intelligent analysis on the time series load of the power load, and then obtains the power load forecasting result.

[0015] The first object of the present invention is to provide a power load forecasting system based on the EFformer model, comprising:

[0016] A human-computer interaction module is used to receive or store the time series load of power load data and output power load forecast results;

[0017] The EFformer model processes the time series load to obtain the power load forecast result; the EFformer model includes an encoder and a decoder; following the signal processing flow:

[0018] The encoder includes an encoder time series load denoising module for processing the time series load, an encoder primary multi-subband decomposition module, an encoder attention module, an encoder forward propagation module, and an encoder secondary multi-subband decomposition module;

[0019] The decoder comprises a decoder time series load denoising module, a decoder primary multi-subband decomposition module, a decoder primary attention module, a decoder secondary multi-subband decomposition module, a decoder secondary attention module and a decoder forward propagation module for sequentially processing the time series load; the decoder primary attention module receives the output result of the encoder secondary multi-subband decomposition module; wherein:

[0020] The encoder time series load denoising module and the decoder time series load denoising module divide the time series load into four-level wavelet subbands, use the high-frequency information of the separated four-level wavelet subbands as noise factors, reconstruct the remaining primary, secondary and tertiary wavelet subbands, and obtain the denoised time series load;

[0021] The encoder primary multi-subband decomposition module, the encoder secondary multi-subband decomposition module, the decoder primary multi-subband decomposition module and the decoder secondary multi-subband decomposition module decompose the denoised time series load into a low-frequency part and a high-frequency part to obtain two load sequences in different frequency domains, the low-frequency part corresponds to the trend feature, and the high-frequency part corresponds to the periodic feature;

[0022] The encoder attention module, the decoder primary attention module and the decoder secondary attention module perform feature extraction on the load sequence, convert the time domain signal into a frequency domain signal, and extract the frequency and amplitude of the load sequence;

[0023] The encoder forward propagation module and the decoder forward propagation module first map the time series load to a high-dimensional space and then map it to a low-dimensional space through linear transformation to extract deeper features;

[0024] The decoder forward propagation module sends the deeper features as the power load prediction results to the human-computer interaction module.

[0025] Preferably, the encoder time series load denoising module and the decoder time series load denoising module adopt a wavelet multi-level decomposition and reconstruction method to divide the time series load into four levels of wavelet sub-bands.

[0026] Preferably, the encoder primary multi-subband decomposition module, the encoder secondary multi-subband decomposition module, the decoder primary multi-subband decomposition module and the decoder secondary multi-subband decomposition module decompose the denoised time series load into a low-frequency part and a high-frequency part using discrete wavelets.

[0027] Preferably, the encoder attention module, the decoder primary attention module and the decoder secondary attention module use fast Fourier to perform feature extraction.

[0028] Preferably, the method for constructing the EFformer model includes:

[0029] S101, obtaining a historical data set related to a power grid; each data in the data set includes M-dimensional features, where M is a natural number greater than 0;

[0030] S102, data set preprocessing; specifically including processing missing values, filtering out maximum values ​​or minimum values;

[0031] S103, model training;

[0032] First, the preprocessed data set is divided into training set, validation set, and test set;

[0033] Then set the Epochs parameter and finally start training the model. When training the model, use the early termination method to prevent overfitting.

[0034] S104, using mean square error and / or mean absolute error to measure the prediction ability of the model for the data in the test set, and generating a prediction data curve graph to judge the accuracy of EFformer prediction.

[0035] Preferably, the data in the data set are transformer data; the M-dimensional features include data point recording date, predicted target oil temperature, high effective load, high invalid load, medium effective load, medium invalid load, low effective load and low invalid load; the power load prediction result is the future oil temperature change of the transformer.

[0036] Preferably, the data set preprocessing includes: processing missing values ​​by data smoothing, and filtering out maximum values ​​or minimum values.

[0037] Preferably, the ratio of the training set, validation set and test set is 7:1:2.

[0038] The second object of the present invention is to provide a method for power load forecasting based on the EFformer model, using the above-mentioned power load forecasting system based on the EFformer model to perform the following steps:

[0039] S1. Obtain the time series load of power load data using the human-computer interaction module;

[0040] S2, importing the time series load into the EFformer model;

[0041] S3, the EFformer model performs data processing on the time series load to obtain a power load forecast result;

[0042] S4. Using a human-computer interaction module to read the power load forecast result.

[0043] Preferably, the encoder first-level multi-subband decomposition module, the encoder second-level multi-subband decomposition module, the decoder first-level multi-subband decomposition module and the decoder second-level multi-subband decomposition module decompose the denoised time series load into a low-frequency part and a high-frequency part, and the number of decompositions is P times, where P is a natural number greater than 0.

[0044] The advantages and positive effects of the present invention are:

[0045] The present invention uses the EFformer model to perform intelligent analysis on the time series load of the power load, and can quickly and accurately obtain the power load forecast results. Specifically:

[0046] The present invention uses wavelet transform to denoise the time series load, decomposing the time series load into different wavelet sub-bands, each of which represents a frequency within a certain range. Among these sub-bands, high-frequency sub-bands usually represent noise and detail information, while low-frequency sub-bands usually represent the overall trend and long-term changes of the signal. By removing the noise in the high-frequency sub-bands, we can better retain the main features of the signal and reduce the errors introduced in signal reconstruction.

[0047] The present invention designs an EF multi-subband decomposition module and an EF attention module. The EF multi-subband decomposition module uses discrete wavelet multi-level decomposition to decompose the denoised time series load into four sub-bands. Each sub-band corresponds to a different frequency domain of the time series load. The EF attention module is responsible for extracting features of the time series loads in different frequency domains respectively and accumulating the prediction results of different sub-bands.

[0048] This paper proposes a new EFformer model based on Transformer, which uses wavelet transform for time series load denoising, and uses EF attention module for feature extraction, and adopts Transformer-based Encoder-Decoder structure to cope with the ability of simultaneously modeling long-term and short-term time series features. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 It is a structural diagram of the EFformer model in the preferred embodiment of the present invention;

[0050] Figure 2 This is a diagram showing the effect of experimental prediction accuracy in a preferred embodiment of the present invention. DETAILED DESCRIPTION

[0051] In order to further understand the content, features and effects of the present invention, the following embodiments are listed and described in detail with reference to the accompanying drawings:

[0052] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the technical solutions in the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0053] Table 1 is an explanation of the English abbreviations

[0054]

[0055] Introduction to the EFformer Model

[0056] The EFformer model is a time series prediction model based on the Transformer structure, which has good adaptability and generalization ability. By encoding and decoding the power load data, the EFformer model can achieve efficient load prediction while retaining the data characteristics.

[0057] The construction process of the EFformer model includes:

[0058] (1) Data preprocessing: Clean, normalize and divide the collected power load data to facilitate subsequent training and prediction.

[0059] (2) Constructing the EFformer model: According to the characteristics of the power load data, design a suitable EFformer model structure and adjust the parameters.

[0060] (3) Training model: Use the divided training data set to train the EFformer model and optimize the model parameters.

[0061] (4) Model evaluation and optimization: By dividing the validation data set, the performance of the model on the load forecasting task is evaluated. Based on the evaluation results, the model parameters are further adjusted to improve the prediction accuracy.

[0062] (5) Practical application: The trained EFformer model is applied to actual power load forecasting to provide strong support for power grid operation management and dispatching.

[0063] The present invention designs and develops a power load forecast based on the EFformer model, which can more accurately forecast the power load and provide a scientific basis for power grid operation management and dispatching. This will help to implement peak shifting and valley filling, improve load rate, reduce the maximum power demand, thereby improving power generation efficiency and optimizing energy allocation. With the continuous development of smart grid technology, more efficient and environmentally friendly energy utilization is expected to be achieved in the future.

[0064] First embodiment:

[0065] An electric load forecasting system based on an EFformer model comprises:

[0066] The human-computer interaction module is used to receive or store the time series load of the power load data on the one hand, and to output the power load forecast result on the other hand;

[0067] The EFformer model processes the time series load to obtain the power load forecast result; please refer to Figure 1 , the EFformer model includes an encoder and a decoder; following the signal processing flow:

[0068] The encoder includes an encoder time series load denoising module for processing the time series load, an encoder primary multi-subband decomposition module, an encoder attention module, an encoder forward propagation module, and an encoder secondary multi-subband decomposition module;

[0069] The decoder comprises a decoder time series load denoising module, a decoder primary multi-subband decomposition module, a decoder primary attention module, a decoder secondary multi-subband decomposition module, a decoder secondary attention module and a decoder forward propagation module for sequentially processing the time series load; the decoder primary attention module receives the output result of the encoder secondary multi-subband decomposition module; wherein:

[0070] The encoder time series load denoising module and the decoder time series load denoising module divide the time series load into four-level wavelet subbands, use the high-frequency information of the separated four-level wavelet subbands as noise factors, reconstruct the remaining low-frequency low-level wavelet subbands, and obtain the denoised time series load;

[0071] The encoder primary multi-subband decomposition module, the encoder secondary multi-subband decomposition module, the decoder primary multi-subband decomposition module and the decoder secondary multi-subband decomposition module decompose the denoised time series load into a low-frequency part and a high-frequency part to obtain two load sequences in different frequency domains, the low-frequency part corresponds to the trend feature, and the high-frequency part corresponds to the periodic feature;

[0072] The encoder attention module, the decoder primary attention module and the decoder secondary attention module perform feature extraction on the load sequence, convert the time domain signal into a frequency domain signal, and extract the frequency and amplitude of the load sequence;

[0073] The encoder forward propagation module and the decoder forward propagation module first map the time series load to a high-dimensional space and then map it to a low-dimensional space through linear transformation to extract deeper features;

[0074] The decoder forward propagation module sends the deeper features as the power load prediction results to the human-computer interaction module.

[0075] The method for constructing the EFformer model includes:

[0076] S101, obtaining a historical data set related to a power grid; each data in the data set includes M-dimensional features, where M is a natural number greater than 0;

[0077] For example: The present invention obtains the oil temperature data of power transformers, which is provided by the State Grid. The data set is transformer data from two different counties in the same province in China, with a time span of 2 years. The original data is recorded once a minute (marked with m), and each data set contains 2 years * 365 days * 24 hours * 60 minutes = 1,051,200 data points. Due to the huge amount of data, the data granularity is changed to 1 data point every 15 minutes, recorded as ETTm1 and ETTm2. Each data point contains 8-dimensional features, including the data point recording date, the predicted target value oil temperature (OT) and six different types of power load characteristics; the predicted target value oil temperature (OT) is the actual recorded data of the same county in China, with a time span of two years, and the six different types of power load characteristics are:

[0078] High Useful Load High Useful Load;

[0079] High Useless Load High Useless Load;

[0080] middle Useful Load Payload;

[0081] Invalid load in middle userless Load;

[0082] low Useful Load low useful load;

[0083] low userless Load;

[0084] The above six parameters all affect the final oil temperature, so they are considered when training the model.

[0085] S102, data set preprocessing; specifically, processing missing values ​​by data smoothing, filtering out maximum or minimum values; preprocessing mainly includes:

[0086] Process missing values, such as those caused by sensor failure or sample data loss, to ensure that there are no null values ​​in the data set;

[0087] Outlier processing is to smooth out the outliers caused by sensor failure or human error to ensure that there are no abnormal maximum or minimum values ​​in the data set.

[0088] S103, model training;

[0089] First, the preprocessed data set is divided into training set, validation set, and test set;

[0090] Then set the Epochs parameter and finally start training the model. When training the model, use the early termination method to prevent overfitting; specifically:

[0091] The cleaned data set is used as input, and the data set is divided into training set, validation set, and test set; the ratio of training set, validation set, and test set is 70%, 10%, and 20% respectively. The training set is used to train the model, the validation set is used to correct and strengthen the learned knowledge, verify whether it is overfitting, and adjust the training parameters, etc. The test set is used to evaluate the performance of the final model. Epochs (the number of times the training algorithm traverses the entire training data set) is set to 10 times (the Epochs value can be adjusted according to the specific situation), and the early termination method is used to prevent overfitting, that is, stop the iteration before the model converges on the training data set to prevent overfitting. The mean square error (MSE) and mean absolute error (MAE) are used to measure the model's prediction ability for the data in the test set, and a prediction data curve is generated to intuitively judge the accuracy of EFformer's prediction.

[0092] S104, using mean square error and / or mean absolute error to measure the prediction ability of the model for the data in the test set, and generating a prediction data curve graph to judge the accuracy of EFformer prediction.

[0093] For example, the data set can be divided into training set, validation set, and test set, with the ratio of training set, validation set, and test set being 7:1:2. The first 70% of the data is used for data training, and 10% of the data is used for verification during training to prevent overfitting. The predicted results are compared with the final 20% of the data to verify the prediction accuracy of the model.

[0094] The power load prediction result is the future oil temperature change of the transformer.

[0095] The EFformer prediction in the present invention is described in detail below:

[0096] The main structure of EFformer prediction adopts encoder + decoder structure;

[0097] The main task of the encoder is to convert the input sequence into a fixed-length context vector, which is a compressed representation of the entire input sequence and captures the main features of the sequence.

[0098] The decoder uses this vector to generate the output sequence. By reintroducing information from the input sequence into the decoding phase, it helps the decoder better utilize detailed information from historical data when generating predictions.

[0099] In some cases, relying solely on the context vector generated by the encoder may not be enough to capture all the necessary information, or this information may be lost during the conversion process. By directly passing the input sequence to the decoder, additional information can be provided during the decoding process to assist the decoder in more accurately reconstructing or predicting future sequences. It includes four sub-modules: Time-Series Denoising Block, EF-Decomp Block, EF-wavelet Attention Block, and FeedForward Block. The model structure is as follows: Figure 1 shown.

[0100] The time-series load denoising module (Time-Series Denoising Block) adopts the wavelet multi-level decomposition and reconstruction method to divide the time series load into four-level wavelet sub-bands, uses the high-frequency information of the separated four-level wavelet sub-bands as the noise factor, and reconstructs the remaining low-frequency and low-level wavelet sub-bands to obtain the denoised time series load.

[0101] The EF-Decomp Block is used to decompose the denoised time series load into a low-frequency part and a high-frequency part using discrete wavelets. Two load sequences in different frequency domains can be obtained. The low-frequency part corresponds to trend characteristics, and the high-frequency part corresponds to periodic characteristics. Moreover, this decomposition is not performed only once, but in a repeated decomposition mode.

[0102] The restored load sequence is sent to the EF-wavelet Attention Block, which uses fast Fourier transform to extract features, convert the time domain signal into frequency domain signal, and extract the frequency and amplitude of the load sequence.

[0103] The feed forward block maps the time series load to a high-dimensional space and then to a low-dimensional space through linear transformation, thereby extracting deeper features.

[0104] Time-Series Denoising Block

[0105] The principle of wavelet transform denoising is based on the sparsity of the signal, that is, the high-frequency components of the signal in the wavelet domain are less than the low-frequency components. Therefore, setting the high-frequency components below the threshold to zero will not have much impact on the signal, but it can effectively remove the noise. This module uses wavelet transform to decompose the time series load into frequency components of three scales, and then thresholds the high-frequency components on each scale, sets the high-frequency components below the threshold to zero, and finally inverse transforms the processed time series load to obtain the denoised time series load. The basic steps are as follows:

[0106] Step 1: First, the signal x(t) is decomposed into wavelet coefficients of different scales by discrete wavelet transform. In actual operation, the present invention uses a series of high-pass and low-pass filter combinations to perform four-layer continuous wavelet decomposition on the original signal x(t). The detailed information (high-frequency part) and approximate information (low-frequency part) of the signal are extracted respectively. Then it can be expressed as:

[0107] First layer:

[0108] = (1);

[0109] = (2);

[0110] Second layer:

[0111] = (3);

[0112] = (4);

[0113] Third layer:

[0114] = (5);

[0115] = (6);

[0116] Fourth floor:

[0117] = (7);

[0118] = (8);

[0119] In this decomposition process, h[k] and g[k] are the coefficients of the low-pass and high-pass filters, respectively, which are part of the specific wavelet basis. represents the low-frequency part of the rth layer, Represents the high frequency part of the rth layer, n is used to represent the time index of the signal or the position in the discrete sequence, and k is used to represent the index of the filter coefficient, that is, the signal is shifted by different amounts, which allows the filter to capture different time characteristics of the signal. The present invention selects Haar wavelet as the wavelet basis. Each decomposition takes the low frequency part as the input of the next level of decomposition.

[0120] The final result of the decomposition is a set of series data, including the low-frequency coefficients of the last layer And the high frequency coefficients of all layers , , , ;

[0121] Step 2: Select a universal threshold λ to determine which wavelet coefficients may represent only noise. The formula is as follows:

[0122] (9);

[0123] Where σ is the noise standard deviation and N is the number of data points

[0124] For the high frequency coefficients of each layer, the present invention uses a hard threshold method for processing, and the formula is as follows:

[0125] [k] = (10);

[0126] Represents the kth coefficient of the high-frequency coefficients after wavelet transformation of the jth layer, [k] is the result after threshold λ screening. This coefficient represents the high-frequency component in the data in wavelet transform. For each layer of high-frequency coefficients , , , Apply the selected threshold as described above. It usually remains unchanged because it contains most of the energy and information of the signal.

[0127] Step 3: Use the processed wavelet coefficients (including high-frequency and low-frequency parts) to reconstruct the original signal.

[0128] From the low frequency part of the lowest frequency layer (the fourth layer) Start by using the processed high-frequency part layer by layer [k] and low frequency part To reconstruct the signal of the previous layer. For each layer j=4,3,2,1;

[0129] = + ](11);

[0130] in, are the inverse filter coefficients corresponding to the low-pass and high-pass filters, respectively, and j represents the number of layers. After four inverse iterative reconstructions, you will get the reconstructed signal (0), which is the denoised signal.

[0131] EF multi-subband decomposition module (EF-Decomp Block);

[0132] The EF multi-subband decomposition module uses the decomposition and reconstruction principle of discrete single-layer wave transform to decompose the time series load into two parts: high-frequency load and low-frequency load. The high-frequency part corresponds to the periodic characteristics of the time series, and the low-frequency load corresponds to the trend characteristics of the time series, which is used to predict the periodic fluctuations and long-term trends of the data.

[0133] For the denoised time series (t), where t is the time index. Our goal is to decompose this time series into high-frequency parts (periodic characteristics) and low-frequency parts (trend characteristics). This experiment selects Haar wavelet and the corresponding low-pass filter h[n] and high-pass filter g[n] to analyze the time series. (0) Applied to extract trends and detail components.

[0134] Low frequency part (trend characteristics):

[0135] [n] = (12);

[0136] here, [n] represents the low-frequency part, which is obtained by applying a low-pass filter to the original sequence and downsampling it. n is used to represent the time index of the signal or the position in the discrete sequence , 𝑘 is used to represent the index of the filter coefficient ,Right now The signal is shifted by different amounts, which allows the filter to capture different temporal characteristics of the signal.

[0137] High frequency part (periodic characteristics):

[0138] [n] = (13);

[0139] here [n] represents the high frequency part, which is obtained by applying a high-pass filter to the original sequence and downsampling it.

[0140] get [n] and [n] respectively reflect the trend and periodic characteristics of the time series. These components can be used for further analysis to predict the long-term trend and periodic fluctuations of the time series.

[0141] EF-wavelet Attention Block;

[0142] The EF attention module takes the high-frequency and low-frequency signals decomposed by the multi-subband decomposition module as input, and uses the EF attention module as a substitute for the self-attention mechanism. The fast Fourier transform (FFT) is used to extract the features of the time series load, and the feature enhancement is repeated three times so that the model can better learn the implicit association of the load.

[0143] The present invention extracts the low-frequency signal and the high-frequency signal from the EF multi-subband decomposition module. [n] and [n] Use fast Fourier transform (FFT) to extract features from the two signals. The signal features are strengthened by repeatedly applying the features extracted by FFT. After each feature extraction, a nonlinear activation function ReLU function is used for processing. The formula is as follows:

[0144] (w) = ReLU(FFT( [n]))(14);

[0145] (w) = ReLU(FFT( [n]))(15);

[0146] The frequency domain data is then converted back to the time domain through inverse Fourier transform, and the obtained periodic signals and trend signals are accumulated to finally obtain the predicted time series.

[0147] A method for predicting electric load based on the EFformer model, using the above-mentioned electric load prediction system based on the EFformer model, performs the following steps:

[0148] S1. Obtain the time series load of power load data using the human-computer interaction module;

[0149] S2, importing the time series load into the EFformer model;

[0150] S3, the EFformer model performs data processing on the time series load to obtain a power load forecast result;

[0151] S4. Using a human-computer interaction module to read the power load forecast result.

[0152] The multi-subband decomposition module decomposes the denoised time series load into a low-frequency part and a high-frequency part, and the number of decompositions is P times, where P is a natural number greater than 0.

[0153] Experimental steps and results

[0154] The data set selected by the present invention is transformer data in real scenarios, provided by the State Grid, with 1 data point every 15 minutes, and each data point contains 8-dimensional features, including the data point recording date, the predicted target value oil temperature (OT) and 6 different types of power load features. This experiment uses the L2 regularization loss function for training, the initial learning rate is set to 0.0001, and the batch size is set to 32. The training process is controlled to stop early within 10 epochs. The model contains 2 encoder layers and 1 decoder layer, and all experiments are repeated four times to ensure the accuracy of the experimental results. The training set / validation set / test set / division ratio is 0.7 / 0.1 / 0.2, the experimental encoder input length is 96, the decoder input length is 48, and the prediction step length is 96 / 192 / 336 / 720 respectively. Transformer, Informer, Autoformer, and Fedformer are selected as baselines, and MAE and MSE are used as the standards for feedback experimental model prediction effects, and the experimental prediction accuracy effect diagram is displayed. The experimental results are as follows:

[0155] Table 2 shows the long time series prediction results

[0156]

[0157] From Table 2 and Figure 2 , we can observe that:

[0158] 1. Compared with the most advanced FEDformer model, the MAE of the EFformer model on the ETTm1 dataset has been improved by an average of 5.5% and the MSE has been improved by an average of 3.7%.

[0159] 2. Compared with the most advanced FEDformer model, the EFformer model performs slightly worse on the ETTm2 dataset. EFformer is more suitable for dealing with real scenarios where the dataset periodicity is poor and there are many noise factors.

[0160] 3. In the experiments with prediction steps of 96, 192, and 336, the EFformer model can achieve better results than other mainstream models.

[0161] The EFformer model proposed in this invention uses wavelet transform to replace the traditional average moving algorithm to denoise the time series load. Due to its good local properties, it can well capture the instantaneous changes and non-stationarity of the signal. It has a very good performance when processing some time series loads with obvious instantaneous changes. It uses a single-layer discrete wavelet decomposition to construct the EF multi-subband decomposition module to separate the periodicity and trend of the time series load, and uses fast Fourier transform for feature extraction multiple times. It discovers dependencies and aggregates information at different load information levels to capture the implicit sequence characteristics in the real time series.

[0162] The present invention can help power companies to rationally plan their power supply. By predicting the power load, computing companies can prepare for power production and allocation in advance to meet the peak power demand that may occur in the future, avoid insufficient or wasteful power supply, and thus improve the quality and efficiency of power supply. Secondly, it is conducive to grid dispatching and operation. Power load forecasting can provide reliable reference data for grid dispatchers, help them formulate scientific and reasonable grid dispatching plans, optimize the operating status of the grid, and improve the stability and reliability of the grid. At the same time, it helps to improve energy utilization efficiency. Power load forecasting can help users rationally plan their electricity consumption and avoid wasting electricity during peak periods of electricity consumption, thereby improving energy utilization efficiency and reducing energy consumption and costs.

[0163] In the above embodiments, it can be implemented in whole or in part by software, hardware, firmware or any combination thereof. When the use is implemented in whole or in part in the form of a computer program product, the computer program product includes one or more computer instructions. When the computer program instructions are loaded or executed on a computer, the process or function described in the embodiment of the present invention is generated in whole or in part. The computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions may be stored in a computer-readable storage medium, or transmitted from one computer-readable storage medium to another computer-readable storage medium. For example, the computer instructions may be transmitted from one website site, computer, server or data center by wired (e.g., coaxial cable, optical fiber, digital subscriber line (DSL) or wireless (e.g., infrared, wireless, microwave, etc.) mode) to another website site, computer, server or data center. The computer-readable storage medium may be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more available media integrated. The available medium may be a magnetic medium (e.g., a floppy disk, a hard disk, a tape), an optical medium (e.g., a DVD), or a semiconductor medium (e.g., a solid-state hard disk SolidState Disk (SSD)), etc.

[0164] The above description is only a preferred embodiment of the present invention and does not limit the present invention in any form. Any simple modification, equivalent changes and modifications made to the above embodiments based on the technical essence of the present invention are within the scope of the technical solution of the present invention.

Claims

1. A power load forecasting system based on the EFformer model, characterized in that: include: A human-computer interaction module is used to receive or store the time series load of power load data and output power load forecast results; The EFformer model processes the time series load to obtain the power load forecast result; the EFformer model includes an encoder and a decoder; following the signal processing flow: The encoder includes an encoder time series load denoising module for processing the time series load, an encoder primary multi-subband decomposition module, an encoder attention module, an encoder forward propagation module, and an encoder secondary multi-subband decomposition module; The decoder comprises a decoder time series load denoising module, a decoder primary multi-subband decomposition module, a decoder primary attention module, a decoder secondary multi-subband decomposition module, a decoder secondary attention module and a decoder forward propagation module for sequentially processing the time series load; The decoder's primary attention module receives the output result of the encoder's secondary multi-subband decomposition module; wherein: The encoder time series load denoising module and the decoder time series load denoising module divide the time series load into four-level wavelet subbands, use the high-frequency information of the separated four-level wavelet subbands as noise factors, reconstruct the remaining primary, secondary and tertiary wavelet subbands, and obtain the denoised time series load; The encoder primary multi-subband decomposition module, the encoder secondary multi-subband decomposition module, the decoder primary multi-subband decomposition module and the decoder secondary multi-subband decomposition module decompose the denoised time series load into a low-frequency part and a high-frequency part to obtain two load sequences in different frequency domains, the low-frequency part corresponds to the trend feature, and the high-frequency part corresponds to the periodic feature; The encoder attention module, the decoder primary attention module and the decoder secondary attention module perform feature extraction on the load sequence, convert the time domain signal into a frequency domain signal, and extract the frequency and amplitude of the load sequence; The encoder forward propagation module and the decoder forward propagation module first map the time series load to a high-dimensional space and then map it to a low-dimensional space through linear transformation to extract deeper features; The decoder forward propagation module sends the deeper features as power load prediction results to the human-computer interaction module; The encoder time series load denoising module and the decoder time series load denoising module adopt a wavelet multi-level decomposition and reconstruction method to divide the time series load into four levels of wavelet sub-bands; The method for constructing the EFformer model includes: S101, obtaining a historical data set related to a power grid; each data in the data set includes M-dimensional features, where M is a natural number greater than 0; S102, data set preprocessing; specifically including processing missing values, filtering out maximum values ​​or minimum values; S103, model training: first, dividing the preprocessed data set into a training set, a validation set, and a test set; Then set the Epochs parameter and finally start training the model. When training the model, use the early termination method to prevent overfitting. S104, using mean square error and / or mean absolute error to measure the prediction ability of the model for the data in the test set, and generating a prediction data curve graph to judge the accuracy of EFformer prediction.

2. The power load forecasting system based on the EFformer model according to claim 1 is characterized in that: The encoder primary multi-subband decomposition module, the encoder secondary multi-subband decomposition module, the decoder primary multi-subband decomposition module and the decoder secondary multi-subband decomposition module decompose the denoised time series load into a low-frequency part and a high-frequency part using discrete wavelets.

3. The power load forecasting system based on the EFformer model according to claim 1 is characterized in that: The encoder attention module, the decoder primary attention module and the decoder secondary attention module use fast Fourier to perform feature extraction.

4. The power load forecasting system based on the EFformer model according to claim 1 is characterized in that: The data in the data set are transformer data; the M-dimensional features include data point recording date, predicted target oil temperature, high effective load, high invalid load, medium effective load, medium invalid load, low effective load and low invalid load; the power load prediction result is the future oil temperature change of the transformer.

5. The power load forecasting system based on the EFformer model according to claim 1 is characterized in that: The data set preprocessing includes: processing missing values ​​through data smoothing, and filtering out maximum values ​​or minimum values.

6. The power load forecasting system based on the EFformer model according to claim 1 is characterized in that: The ratio of the training set, validation set, and test set is 7:1:

2.

7. A method for power load forecasting based on the EFformer model, characterized in that: Using the power load forecasting system based on the EFformer model as described in any one of claims 1 to 6, the following steps are performed: S1. Obtain the time series load of power load data using the human-computer interaction module; S2, importing the time series load into the EFformer model; S3, the EFformer model performs data processing on the time series load to obtain a power load forecast result; S4. Using a human-computer interaction module to read the power load forecast result.

8. The power load forecasting method based on the EFformer model according to claim 7 is characterized in that: The encoder primary multi-subband decomposition module, the encoder secondary multi-subband decomposition module, the decoder primary multi-subband decomposition module and the decoder secondary multi-subband decomposition module decompose the denoised time series load into a low-frequency part and a high-frequency part, the number of decompositions is P times, and P is a natural number greater than 0.