Power load prediction method and system based on frequency domain adaptive filter

Abstract

The invention discloses a power load prediction method and system based on a frequency domain adaptive filter. The method comprises the following steps: cleaning power load related data to generate available time sequence data; performing discrete Fourier transform on the time sequence data to convert time domain data into frequency domain data; introducing a frequency domain adaptive normalization layer, performing amplitude filtering on frequency domain data, distinguishing stationary frequency domain data and non-stationary frequency domain data, and recovering time domain data by adopting inverse discrete Fourier transform; performing normalization processing on the stationary time domain data and the non-stationary time domain data respectively; extracting features by using the frequency domain adaptive prediction model, and generating corresponding stationary and non-stationary time domain prediction data; and performing weighted fusion on the results of the stationary time domain prediction data and the non-stationary time domain prediction data to generate complete time sequence prediction data. According to the method, the prediction capability of the stationary time domain data is remarkably improved, and meanwhile, the change rule in the non-stationary data can be more accurately captured, so that the overall prediction precision is enhanced.

Description

Technical Field

[0001] The present invention belongs to the technical field of power system operation and dispatch, and particularly relates to a power load forecasting method and system based on a frequency-domain adaptive filter. Background Art

[0002] Power load forecasting is the basis for the optimal dispatch and safe operation of power systems, directly affecting the formulation of power supply plans, the allocation of energy resources, and demand response strategies. However, power load data has complex characteristics such as non-linearity, volatility, and intermittency, and contains obvious stationary and non-stationary information. Traditional power load forecasting methods often fail to effectively separate and model these complex characteristics when dealing with such data, resulting in deficiencies in the accuracy and robustness of the models. Existing load forecasting models generally face the following problems: First, the multi-scale characteristics of power load data make it difficult for traditional models to effectively fuse different time dimensions and data features; second, due to the non-stationarity of load data, traditional models are difficult to capture the dynamic changes in the data, especially in the relationship between high-frequency fluctuations and low-frequency trends; finally, the multi-variable characteristics in power load data often intertwine with each other, making it difficult for the model to effectively extract the correlations between variables.

[0003] Existing time series modeling methods, such as regression models and support vector machines, although having certain forecasting capabilities, usually assume that the load data is stationary and linear, and are difficult to cope with complex non-stationary characteristics and frequency-domain features. Deep learning models such as long short-term memory networks (LSTMs) and Transformers, although having significant advantages in non-linear modeling, still have limitations in dealing with non-stationarity, noise, and frequency-domain feature extraction. In particular, the non-stationarity, periodic fluctuations, and random noise commonly existing in power load data will seriously affect the performance of the forecasting model. Therefore, there is an urgent need for a forecasting model that can effectively separate stationary and non-stationary characteristics and make full use of frequency-domain information. Summary of the Invention

[0004] To solve the deficiencies of the prior art, the present invention combines frequency-domain adaptive filtering technology with time series modeling methods to achieve the purpose of effectively separating and processing non-stationary characteristics in data and improving the accuracy and robustness of power load forecasting. The present invention adopts the following technical solutions:

[0005] A power load forecasting method based on a frequency-domain adaptive filter includes the following steps:

[0006] Step 1: Obtain the time series data of the power load from the relevant data of the power load;

[0007] In Step 1, the power load-related data includes historical power load data, meteorological information, special event information, and user behavior characteristic data. The user behavior characteristic data includes the electricity consumption characteristics of different types of users. Clean the power load-related data, including handling missing values and outliers, and perform smoothing processing by the moving average method to remove random noise. Then, use the smoothed data as the input time series data. The formula for the smoothing process is as follows:

[0008]

[0009] where y t represents the smoothed data, x i represents the original data point, N represents the window size, and t represents time.

[0010] Step 2: Perform a time-domain to frequency-domain conversion on the time series data to obtain a frequency-domain signal;

[0011] In Step 2, use the discrete Fourier transform for the time-domain to frequency-domain conversion, which can greatly reduce the computational complexity. Suppose there is a time-domain signal x(t) = {x0, x1, …, x N-1} of length N. The time series data is transformed into the corresponding frequency-domain signal X(f) through the discrete Fourier transform algorithm. The formula is:

[0012]

[0013] where X(f) represents the frequency-domain signal, which contains the amplitude and phase information of different frequency components in the signal.

[0014] Step 3: Introduce a frequency-domain adaptive normalization layer (L-FAN) to perform amplitude filtering on the frequency-domain signal, distinguish the frequency-domain signal into stationary frequency-domain data and non-stationary frequency-domain data, then convert the frequency-domain data back to time-domain data, and perform normalization processing on the stationary time-domain data and non-stationary time-domain data respectively;

[0015] Step 4: Construct and utilize a frequency-domain adaptive prediction model to respectively perform feature extraction on the normalized stationary time-domain data and non-stationary time-domain data and generate corresponding stationary time-domain prediction data and non-stationary time-domain prediction data;

[0016] Step 5: Fuse the stationary time-domain prediction data and the non-stationary time-domain prediction data to generate a complete time series target output. The formula is as follows:

[0017] Y final = α · Y stationary + (1 - α) · Y non-stationary

[0018] Among them, α represents the weight factor, which is usually determined according to the quality or prediction accuracy of stationary time domain data and non-stationary time domain data.

[0019] Furthermore, in step 3, the amplitude of each frequency domain component is first extracted from the frequency domain signal, and the frequency domain component with a large amplitude in the low-frequency part is used as stationary frequency domain data, and the frequency domain component with a small amplitude in the high-frequency part is used as non-stationary frequency domain data.

[0020] The frequency domain signal is a complex frequency domain signal, and the formula is as follows:

[0021]

[0022] Among them, |X(f)| represents the amplitude of the frequency domain component extracted from the frequency domain signal X(f), Re(X(f)) and Im(X(f)) represent the real part and imaginary part of the frequency domain signal X(f), respectively.

[0023] In step 3, the frequency domain data is subjected to inverse discrete Fourier transform, and the formula for converting it back to the time domain signal is as follows:

[0024]

[0025] Among them, Y k represents frequency domain data, X n Represents the time domain signal after inverse transformation, k represents the index of the time series corresponding to the frequency domain data, n represents the index of the time series corresponding to the time series signal, k and n are not necessarily equal, because some time domain data may be Fourier transformed into the same frequency domain data, and N represents the length of the time series.

[0026] Furthermore, in step 3, the stationary time domain data is normalized as follows:

[0027]

[0028] Among them, X stable represents stable time domain data, Norm represents normalization operation, Represents the time domain data of the first L time steps of the i-th time domain data point, Mean L Indicates the operation of calculating the average value along the time dimension, Std L Represents the operation of calculating the standard deviation along the time dimension, N represents the length of the stationary time domain data, and L represents the length of the lookback window;

[0029] The non-stationary time domain data is normalized as follows:

[0030]

[0031] For each non-stationary time domain data X unstableThe affine transformation is a linear transformation through the weight parameter affine weight and the bias parameter affine bias to enhance the expressive power of the model and obtain the normalized non-stationary time-domain data σ represents the standard deviation, N′ represents the length of the non-stationary time-domain data, and μ represents the mean.

[0032] Furthermore, in step 4, first, the discrete Fourier transform is performed on the first and second dimensions of the stationary time-domain data, converted to the frequency domain, and orthogonalization processing is performed to ensure that the energy before and after the transformation is consistent. It considers the symmetry of the input sequence during the transformation to avoid redundant frequency-domain components. The formula is as follows:

[0033]

[0034] where X k represents the k-th frequency-domain component in the frequency domain, x n represents the n-th value of the normalized stationary time-domain data N represents the length of the input time-domain data, j represents the imaginary unit, and finally, the time-series data corresponding to the converted frequency domain is obtained L represents the length of the backtracking window, represents the set of time-series data;

[0035] Then, based on the converted time-series data the output of the learnable filter is generated through the learnable filter The formula is as follows: The formula is as follows:

[0036]

[0037] where, represents the Fourier transform, represents the inverse Fourier transform, and ⊙ L represents the element-wise product along the L dimension.

[0038] Furthermore, for the frequency-domain adaptive prediction model in step 4, there are usually obvious stationary and non-stationary signal data in the power load data. However, the linear mapping and attention mapping widely used in existing time-series prediction models have limitations in dealing with stationary frequency-domain data and are difficult to efficiently handle frequency-domain characteristics and non-stationarity problems. The present invention uses a learnable filter module to extract features from the stationary time-domain data, dynamically capture the main frequency-domain feature patterns, and obtain the prediction result Y of the stationary time-domain data stationary ; for the non-stationary time-domain data, a multi-layer perceptron (MLP) is used to model the first K time-domain data, extract the evolution law of the non-stationary frequency-domain pattern, and obtain the prediction result Y of the non-stationary time-domain datanon-stationary , finally, through weighting, the prediction result Y of the stationary time-domain data stationary and the prediction result Y of the non-stationary time-domain data non-stationary are aggregated to obtain the final prediction result Y final , and the formula is as follows:

[0039] Y final = α·Y stationary +(1 - α)·Y non-stationary

[0040] Y stationary = Learning_Filter(X stationary )

[0041] Y non-stationary = MLP(X non-stationary )

[0042] where α represents the weight factor, Learning_Filter represents the learnable filter, and MLP represents the multi-layer perceptron.

[0043] Furthermore, for the learnable filter module, by randomly initializing the learnable parameters and performing a multiplication operation with the input stationary time-domain data, an initial learnable filter is generated; a channel-independent strategy is adopted to model the channels of the stationary time-domain data, where the channel-independent strategy models each channel separately to improve the modeling accuracy of the filter and is used to capture the main features between different channels; the learnable filter dynamically adjusts the weights of the filter for different frequency-domain components according to the frequency-domain features in the time-domain data, and the filter can make adjustments for high-frequency changes and low-frequency trends to extract key frequency-domain features during the modeling process and improve the modeling ability of the stationary time-domain data.

[0044] Furthermore, the output of the learnable filter module, after passing through a feed-forward neural network composed of two layers of linear transformation and a Leaky ReLU activation function, obtains the prediction result of the stationary time-domain data. The formula of the Leaky ReLU activation function is as follows:

[0045]

[0046] When x ≥ 0, the output is x, the same as the standard ReLU; when x < 0, the output is αx, where α represents a small positive number, usually taking a value between 0.01 or 0.1, which is used to control the slope of the negative input part and avoid the "dead neuron" problem where negative values are directly output as 0 in the standard ReLU; the Leaky ReLU activation function allows a small negative slope, enabling the neuron to remain activated even when receiving negative input, thereby improving the learning ability of the model.

[0047] Furthermore, the multi-layer perceptron extracts the feature patterns in non-stationary time-domain data, including multiple fully connected layers, each layer containing a non-linear activation function for mining the deep features in the data. For the frequency features of non-stationary time-domain data, the time-domain components of the first K frequencies are selected, and the multi-layer perceptron is used to model and predict these time-domain components of the frequencies to capture the key change patterns of non-stationary time-domain data.

[0048] Although the influence of non-stationary data on the prediction results is eliminated during the normalization process, in the real world, the non-stationarity in the data often reveals the volatility and intermittency that future data may contain. Therefore, it is necessary to model non-stationary data separately. To effectively model the changing pattern of non-stationary information, a multi-layer perceptron is used to perform a linear transformation on the non-stationary time-domain data, mapping the data from one space to another through matrix operations to achieve feature extraction, dimension adjustment, and data representation. It is an essential tool in neural network modeling and dimensionality reduction, providing a basis for non-linear activation and simplifying large-scale data calculations. The linear transformation formula is as follows:

[0049] y = W·x + b

[0050] where W represents the weight matrix, x represents the input data, b represents the bias term, and y represents the output of the linear transformation. The ReLU activation function is used for the result of the linear transformation, and the formula is: ReLU(x) = max(0, x). This operation can introduce non-linear characteristics for feature extraction to enhance the model's expressive ability.

[0051] Furthermore, the output of the multi-layer perceptron passes through a linear projection layer composed of two layers of linear transformation and a ReLU activation function to map and generate the prediction result from the intermediate prediction result of the multi-layer perceptron, and then through normalization processing, the prediction result of the non-stationary time-domain data is obtained.

[0052] The power load prediction system based on the frequency-domain adaptive filter includes a data acquisition module, a frequency-domain conversion module, a frequency-domain normalization module, a frequency-domain adaptive prediction module, and a weighted fusion module. According to the power load prediction method based on the frequency-domain adaptive filter, the acquisition of time series data of the power load, the conversion and normalization of the frequency-domain signal, the feature extraction and prediction, and the weighted fusion of the prediction data are sequentially performed to generate a complete time series target output.

[0053] The advantages and beneficial effects of the present invention are as follows:

[0054] The present invention significantly improves the processing ability of stationary frequency-domain data by introducing learnable filters to replace traditional linear mapping and attention mapping. In particular, it performs more excellently in capturing high-frequency changes and low-frequency trends, and can more accurately identify high-frequency fluctuations and low-frequency trends in signals, thereby improving the prediction accuracy and adaptability. By introducing a frequency-domain adaptive normalization layer, the present invention can effectively distinguish between stationary and non-stationary data, eliminate the interference of non-stationarity on model prediction, and ensure the adaptability of the prediction model to different data types through an adaptive processing method, significantly improving the accuracy and information integrity of the prediction results. By using a multi-layer perceptron to model non-stationary patterns, especially the optimization of the first K frequency-domain signals, the present invention improves the dynamic response ability of the model to non-stationary data. This mechanism enables the prediction model to more accurately capture the changing rules in non-stationary data, thereby enhancing the overall prediction ability. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 is a flowchart of the method according to an embodiment of the present invention.

[0056] Figure 2 is a framework diagram of a prediction model based on a frequency-domain adaptive filter in an embodiment of the present invention.

[0057] Figure 3 is a schematic structural diagram of a stationary frequency-domain data processing module in an embodiment of the present invention.

[0058] Figure 4 is a schematic structural diagram of a non-stationary frequency-domain data processing module in an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0059] The following further describes the specific embodiments of the present invention in detail with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are only for explaining and illustrating the present invention, and are not intended to limit the present invention.

[0060] The present invention relates to a power load prediction method based on a frequency-domain adaptive filter, which improves the accuracy and robustness of power load prediction and provides support for the optimal scheduling of power systems, as Figure 1 、 Figure 2 shown, and specifically includes the following implementation steps:

[0061] Step 1: Obtain power load-related data and clean the data to obtain available time series data; specifically, the obtained power load-related data is time series data, and missing values and outliers in the data are processed.

[0062] To remove random noise, the power load-related data can be smoothed by a moving average method, and the formula is as follows:

[0063]

[0064] Among them, y t is the smoothed data, x i is the original data point, and N is the window size;

[0065] Then, the smoothed data is used as the input time series data.

[0066] Step 2: Perform frequency domain conversion on the time series data using the discrete Fourier transform to convert the time domain data into frequency domain data, obtaining a frequency domain signal;

[0067] The discrete Fourier transform can greatly reduce the computational complexity. Suppose there is a time domain signal x(t) = {x0, x1, …, x N-1}, and the corresponding frequency domain signal X(f) of the time series data is obtained through the discrete Fourier transform algorithm. The formula is as follows:

[0068]

[0069] Among them, X(f) is the frequency domain signal, containing the amplitude and phase information of different frequency domains in the signal.

[0070] Step 3: Introduce a frequency domain adaptive normalization layer to perform amplitude filtering on the frequency domain signal, distinguish the frequency domain signal into stationary frequency domain data and non-stationary frequency domain data, and use the inverse discrete Fourier transform to convert the frequency domain data into time domain data; and, perform normalization processing on the stationary frequency domain data and the non-stationary frequency domain data respectively, specifically including the following steps:

[0071] First, extract the amplitude of each frequency domain component from the frequency domain signal. Given the frequency domain signal X(f), the amplitude |X(f)| of each frequency domain component can be calculated, where X(f) is a complex frequency domain signal, and the formula is as follows:

[0072]

[0073] Among them, Re(X(f)) and Im(X(f)) are the real part and the imaginary part of the frequency domain signal respectively.

[0074] The key of the frequency domain adaptive normalization layer lies in distinguishing stationary data and non-stationary data, and this distinction is based on the size of the frequency domain or the amplitude of the frequency domain component; in practical applications, a threshold is used to distinguish stationary data and non-stationary data.

[0075] Suppose the amplitude |X(f)| of each frequency domain component, the following criteria can be used to distinguish stationary data and non-stationary data:

[0076] Stationary data: the low-frequency part, usually the frequency-domain components with larger amplitudes.

[0077] Non-stationary data: the high-frequency part, usually the frequency-domain components with smaller amplitudes.

[0078] Specifically, the frequency-domain component f of the stationary data stable will satisfy a certain amplitude threshold condition. Assuming the threshold is T, then the condition for stationary data can be defined as:

[0079] |X(f)| ≥ T

[0080] While the frequency-domain component f of the non-stationary data unstable satisfies:

[0081] |X(f)| < T

[0082] After distinguishing between stationary frequency-domain data and non-stationary frequency-domain data, perform the inverse discrete Fourier transform on all frequency-domain Y k to convert it back to the time-domain signal. The formula is as follows:

[0083]

[0084] where Y k is the frequency-domain signal, and X n is the time-domain signal after the inverse transform.

[0085] Next, normalize these two types of data separately. For the normalization process of stationary data, it is as follows:

[0086] For the time-series input X stable use the instance normalization method, denoted as Norm, which can be expressed as:

[0087]

[0088] where Mean L represents the operation of calculating the average along the time dimension, Std L represents the operation of calculating the standard deviation along the time dimension, N represents the length of the input signal, and L represents the length of the backtracking window.

[0089] For the normalization process of non-stationary data, each data point X unstable will subtract the mean μ, where the mean μ is the mean of the data set X unstable and N′ is the total number of data points. The calculation formula is:

[0090]

[0091] Assume the standard deviation σ is:

[0092]

[0093] Then, the standardization operation divides each data point X stable by the standard deviation σ:

[0094]

[0095] For each data point X unstable , the affine transformation performs a linear transformation through the weights affine weight and the bias affine bias :

[0096]

[0097] Here, affine weight and affine bias are learnable parameters that can be updated through gradients during the training process. This transformation adjusts the scale and offset of the data, enhancing the expressive power of the model.

[0098] Step 4: Construct and utilize a frequency-domain adaptive prediction model to respectively extract features from the normalized stationary time-domain data and non-stationary time-domain data and generate corresponding stationary time-domain prediction data and non-stationary time-domain prediction data, specifically including:

[0099] Use a learnable filter to replace traditional linear mapping and attention mapping for feature extraction of stationary time-domain data, dynamically capture the main time-domain feature patterns, especially in dealing with high-frequency changes and low-frequency trends, and improve the modeling ability of stationary time-domain data; at the same time, for the non-stationary time-domain data separated from the input data, use a multi-layer perceptron for modeling to predict the first K frequency-domain signals and capture the evolution law of non-stationary patterns;

[0100] Considering that there are usually obvious stationary and non-stationary signal data in power load data, but the linear mapping and attention mapping widely used in existing time series prediction models have limitations in dealing with stationary time-domain data and are difficult to efficiently handle time-domain characteristics and non-stationarity problems.

[0101] Aiming at the deficiencies of existing prediction models, the present invention proposes a power load prediction method based on a frequency-domain adaptive filter, using the learnable filter to overcome the above problems, and the specific steps are as follows:

[0102] As Figure 3 shown, first perform a discrete Fourier transform on the first and second dimensions of the stationary time-domain data, convert x to the frequency domain, and perform orthogonalization processing to ensure that the energy before and after the transformation is consistent. It takes into account the symmetry of the input sequence during the transformation, thereby avoiding redundant frequency-domain components. The formula is as follows:

[0103]

[0104] Among them, X k is the k-th frequency domain component in the frequency domain, and x n is the n-th value of the time domain signal of the input signal, N is the length of the input signal, j is the imaginary unit, and finally the converted time series data L is the length of the backtracking window.

[0105] Next, given the input time series and the learnable filter The formula is as follows:

[0106]

[0107] Among them, is the Fourier transform, is the inverse Fourier transform, and ⊙ L represents the element-wise product along the L dimension, represents the channel-independent learnable filter, Ind represents the channel-independent learnable filter strategy, represents the total number of channels, is the output of the learnable filter.

[0108] Based on the above components, a learnable filter prediction model is constructed, and the overall formula is as follows:

[0109] Y stationary = Learning_Filter(X stationary )

[0110] After obtaining the intermediate prediction result, it passes through a feedforward neural network composed of two layers of linear transformation and a Leaky ReLU activation function. The formula of the Leaky ReLU activation function is as follows:

[0111]

[0112] When x ≥ 0, the output is x, which is the same as the standard ReLU; when x < 0, the output is αx, where α is a small positive number, usually taking a value between 0.01 or 0.1, used to control the slope of the negative input part, avoiding the "dead neuron" problem where the negative value is directly output as 0 in the standard ReLU; Leaky ReLU can allow a small negative slope, enabling the neuron to remain activated even when receiving negative input, thus improving the learning ability of the model.

[0113] The intermediate prediction result is mapped to generate the final prediction result. After the final prediction result is normalized, the target output of the stationary prediction data is obtained.

[0114] The core of the learnable filter lies in dynamically adjusting the weights of the filter for different frequency-domain components according to the characteristics of the frequency-domain signal. Specifically, the filter can make adjustments for high-frequency variations and low-frequency trends, extract key frequency-domain features during the modeling process, and improve the modeling ability of stationary time-domain data.

[0115] Although the influence of non-stationary data on the prediction results is eliminated during the normalization process, in the real world, the non-stationarity in the data often reveals the volatility and intermittency that future data may contain. Therefore, it is particularly important to separately model non-stationary data.

[0116] To effectively model the changing patterns of non-stationary information, the present invention proposes a power load forecasting method based on a frequency-domain adaptive filter, and uses the multi-layer perceptron to overcome the above problems. The specific steps are as follows:

[0117] As Figure 4 shown, for non-stationary time-domain data, first, the linear transformation maps the data from one space to another through matrix operations, achieving feature extraction, dimension adjustment, and data representation. It is an indispensable tool in neural network modeling and dimensionality reduction, provides a basis for non-linear activation, and simplifies large-scale data calculations. The linear transformation formula is as follows:

[0118] y = W·x + b

[0119] where W is the weight matrix, x is the input data, b is the bias term, and y is the output.

[0120] Apply the ReLU activation function to the result of the linear transformation. The formula is: ReLU(x) = max(0, x). This operation can introduce non-linear characteristics and enhance the model's expressive ability.

[0121] The overall formula of the multi-layer perceptron prediction model is as follows:

[0122] Y non-stationary = MLP(X non-stationary )

[0123] After obtaining the intermediate prediction result, it passes through a linear projection layer composed of two layers of linear transformation and a ReLU activation function to map the intermediate prediction result to generate the final prediction result. After the final prediction result is normalized, the target output of the non-stationary prediction data is obtained.

[0124] The core of the multi-layer perceptron lies in extracting the feature patterns in non-stationary time-domain data through a multi-layer neural network. The multi-layer perceptron includes multiple fully connected layers, and each layer contains a non-linear activation function for mining the deep features in the data. For the frequency features of non-stationary time-domain data, the first K frequency components are selected, and the multi-layer perceptron is used to model and predict these frequency components to capture the key change patterns of non-stationary data.

[0125] Step 5: Weightedly fuse the results of the stationary time-domain prediction data and the non-stationary time-domain prediction data described in Step 4 to generate a complete time series target output.

[0126] By combining the prediction results of stationary data and non-stationary data through weighting, the interference of non-stationary data on the prediction of stationary frequency-domain data can be reduced. The formula is as follows:

[0127] Y final =α·Y stationary +(1-α)·Y non-stationary

[0128] Where α is the weight factor, which is usually determined according to the quality or prediction accuracy of stationary data and non-stationary data.

[0129] Finally, the model of the present invention can not only effectively handle the non-stationarity in power load data, but also significantly improve the frequency-domain modeling ability of complex time series, and has strong engineering applicability and promotion value.

[0130] A power load prediction system based on a frequency-domain adaptive filter includes: a data acquisition and cleaning module, a frequency-domain conversion module, a frequency-domain normalization module, a feature extraction and prediction module, and a weighted fusion module.

[0131] The data acquisition and cleaning module is used to acquire power load-related data and clean the data to generate available time series data;

[0132] The frequency-domain conversion module is used to perform a discrete Fourier transform on the time series data to convert the time-domain data into frequency-domain data and obtain a frequency-domain signal;

[0133] The frequency-domain normalization module is used to perform amplitude filtering on the frequency-domain signal, distinguish the frequency-domain signal into stationary frequency-domain data and non-stationary frequency-domain data, convert the frequency-domain data back to time-domain data through an inverse discrete Fourier transform, and respectively perform normalization processing on the stationary time-domain data and the non-stationary time-domain data;

[0134] The feature extraction and prediction module is used to extract features from the normalized stationary time-domain data to generate stationary time-domain prediction data; extract features from the normalized non-stationary time-domain data to generate non-stationary time-domain prediction data;

[0135] A weighted fusion module for weighted fusion of the results of stationary time-domain prediction data and non-stationary time-domain prediction data to generate a complete time-series target output.

[0136] On the other hand, the present invention provides a computer-readable storage medium storing program code for execution by a device, the program code including steps for performing any of the implementation methods in the power load prediction method based on a frequency-domain adaptive filter.

[0137] On the other hand, the present invention provides an electronic device including a processor, a memory, and a program or instruction stored on the memory and executable on the processor, the program or instruction implementing steps for performing any of the implementation methods in the power load prediction method based on a frequency-domain adaptive filter when executed by the processor.

[0138] The above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements for some or all of the technical features; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. The power load forecasting method based on frequency domain adaptive filter is characterized by The steps include: Step 1: Obtain time series data of power load from relevant data of power load; Step 2: Convert the time series data from the time domain to the frequency domain to obtain a frequency domain signal; Step 3: Introduce a frequency domain adaptive normalization layer, perform amplitude filtering on the frequency domain signal, divide the frequency domain signal into stationary frequency domain data and non-stationary frequency domain data, then convert the frequency domain data back to time domain data, and perform normalization processing on the stationary time domain data and the non-stationary time domain data respectively; Step 4: construct and use a frequency domain adaptive prediction model to extract features from the normalized stationary time domain data and the non-stationary time domain data, respectively, and generate corresponding stationary time domain prediction data and non-stationary time domain prediction data; Step 5: Fuse the stationary time domain prediction data and the non-stationary time domain prediction data to generate a complete time series target output.

2. The method for predicting power load based on frequency domain adaptive filter according to claim 1, characterized in that: In step 3, the amplitude of each frequency domain component is first extracted from the frequency domain signal, and the frequency domain component with a large amplitude in the low-frequency part is used as stationary frequency domain data, and the frequency domain component with a small amplitude in the high-frequency part is used as non-stationary frequency domain data.

3. The method for predicting power load based on frequency domain adaptive filter according to claim 1, characterized in that: In step 3, the stationary time domain data is normalized as follows: Among them, X stable represents stable time domain data, Norm represents normalization operation, Represents the time domain data of the first L time steps of the i-th time domain data point, Mean L Indicates the operation of calculating the average value along the time dimension, Std L Represents the operation of calculating the standard deviation along the time dimension, N represents the length of the stationary time domain data, and L represents the length of the lookback window; The non-stationary time domain data is normalized as follows: For each non-stationary time domain data X unstable The affine transformation is achieved by weighting the parameters affine weight and bias parameter affine bias Perform linear transformation to obtain normalized non-stationary time domain data σ represents the standard deviation, N′ represents the length of the non-stationary time domain data, and μ represents the mean.

4. The method for power load forecasting based on frequency domain adaptive filter according to claim 1, characterized in that: In step 4, the stationary time domain data is first subjected to discrete Fourier transform, converted to the frequency domain, and orthogonalized. The symmetry of the input sequence is considered during the conversion. The formula is as follows: Among them, X k represents the kth frequency domain component in the frequency domain, x n Represents the normalized stationary time domain data The nth value of , N represents the length of the input time domain data, j represents the imaginary unit, and finally the time series data corresponding to the frequency domain after conversion is obtained L represents the length of the lookback window. Represents a collection of time series data; Then, based on the transformed time series data Through learnable filters Generate output of learnable filter The formula is as follows: in, represents the Fourier transform, represents the inverse Fourier transform, ⊙ L represents the element-wise product along dimension L.

5. The method for predicting power load based on frequency domain adaptive filter according to claim 1, characterized in that: The frequency domain adaptive prediction model in step 4 uses a learnable filter module to extract features from the stationary time domain data, dynamically captures the main frequency domain feature patterns, and obtains the prediction results of the stationary time domain data; for the non-stationary time domain data, a multi-layer perceptron is used to model the first K time domain data, extract the evolution law of the non-stationary frequency domain pattern, and obtain the prediction results of the non-stationary time domain data. Finally, the prediction results of the stationary time domain data and the prediction results of the non-stationary time domain data are aggregated by weighting to obtain the final prediction results.

6. The method for power load forecasting based on frequency domain adaptive filter according to claim 5, characterized in that: The learnable filter module generates an initial learnable filter by randomly initializing learnable parameters and performing multiplication operation with the input stationary time domain data; a channel independence strategy is used to model the channel of the stationary time domain data, wherein the channel independence strategy models each channel separately to capture the main features between different channels; The learnable filter dynamically adjusts the weights of the filter on different frequency domain components according to the frequency domain characteristics in the time domain data to extract key features.

7. The method for power load forecasting based on frequency domain adaptive filter according to claim 5, characterized in that: The output of the learnable filter module is transformed through a feedforward neural network consisting of two layers of linear transformation and a Leaky ReLU activation function to obtain the prediction result of the stationary time domain data. The formula of the Leaky ReLU activation function is as follows: When x ≥ 0, the output is x, which is the same as the standard ReLU; when x < 0, the output is αx, where α represents a small positive number.

8. The method for power load forecasting based on frequency domain adaptive filter according to claim 5, characterized in that: The multi-layer perceptron includes multiple fully connected layers, each layer includes a nonlinear activation function, and is used to mine deep features in the data; for the frequency characteristics of non-stationary time domain data, the time domain components of the first K frequencies are selected, and the time domain components of these frequencies are modeled and predicted using the multi-layer perceptron to capture the key change patterns of the non-stationary time domain data; The multilayer perceptron performs linear transformation on non-stationary time domain data, maps data from one space to another space through matrix operation, so as to realize feature extraction, dimension adjustment and data representation; an activation function is used on the result of the linear transformation to introduce nonlinear characteristics for feature extraction.

9. The method for power load forecasting based on frequency domain adaptive filter according to claim 5, characterized in that: The output of the multilayer perceptron is passed through a linear projection layer consisting of two layers of linear transformation and a ReLU activation function to map the intermediate prediction results of the multilayer perceptron to generate prediction results, and then normalized to obtain the prediction results of non-stationary time domain data.

10. An electric load forecasting system based on a frequency domain adaptive filter comprises a data acquisition module, a frequency domain conversion module, a frequency domain normalization module, a frequency domain adaptive prediction module and a weighted fusion module. According to the electric load forecasting method based on a frequency domain adaptive filter according to any one of claims 1 to 9, the acquisition of time series data of the electric load, conversion and normalization of the frequency domain signal, feature extraction and prediction, and weighted fusion of the predicted data are sequentially performed to generate a complete time series target output.

Images