Long-term power load prediction method based on cross-domain multi-scale self-attention

Abstract

The invention discloses a power load prediction method based on a multi-scale frequency domain state space model. The method comprises the steps of 1, collecting time sequence data and performing preprocessing; 2, constructing a prediction network containing dual-path fast Fourier transform (FFT) and a multi-scale Mangbar module, extracting periodic features through frequency domain decomposition, and processing global and local dependency relationships in combination with a multi-scale architecture so as to obtain a power load prediction result; and 3, training the model by using a loss function until convergence to obtain a power load prediction model. According to the method, the problems of serious noise interference and low feature extraction efficiency of a traditional time domain model in power load long sequence prediction are solved, the prediction precision and the calculation efficiency are improved through frequency domain analysis and multi-scale modeling, and the method is suitable for long-term prediction of a high-dimensional and non-stationary power load time sequence.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of long-term power load prediction, in particular to a long-term time series prediction method combining frequency domain analysis and multi-scale modeling. BACKGROUND

[0002] With the accelerated progress of global industrialization and the vigorous development of new energy industry, the demand for electricity in various fields of society is showing an explosive growth trend. At the same time, the limited nature of traditional power generation resources and the volatility of intermittent renewable energy grid connection have jointly exacerbated the structural contradictions between power supply and demand. Even in areas with relatively complete power grid infrastructure, the fluctuation range of power load in different seasons and different economic cycles is still very significant, especially in the low valley period of electricity consumption, there is a lot of idle power generation capacity, highlighting the severe challenges faced by the power system in dynamic adjustment and resource optimization allocation. The dual dilemma of low utilization rate of power generation equipment and continuous rise of electricity load not only leads to the decline of power supply stability, but also causes waste of energy resources and surge of power grid operation cost, seriously restricting the sustainable development of the power industry. Long-term power load prediction, i.e. systematic analysis and quantitative estimation of the trend of power consumption in the future months or even years, plays a core role in solving the above problems. By accurately grasping the long-term evolution law of power load, power enterprises can scientifically plan power source construction and power grid expansion scheme, reasonably arrange power generation equipment maintenance plan, and effectively coordinate different energy generation proportions. This forward-looking planning strategy not only can improve the operation stability and reliability of the power system, but also can promote the efficient use of energy resources, help achieve the carbon emission reduction target, and reduce the comprehensive operation cost of power enterprises. However, the current mainstream long-term power load prediction method has many limitations. Disturbed by complex external factors such as economic policy adjustment, extreme weather events, and technological innovation, existing models are difficult to fully capture the nonlinear characteristics of power load changes. At the same time, due to the large amount of data samples required for long-term prediction, the difficulty of data cleaning and integration is high, and the model is difficult to achieve an ideal balance between calculation efficiency and prediction accuracy. In addition, most prediction models are not adaptable to the new changes brought by the rapid increase of new energy generation penetration rate, resulting in a large deviation between the prediction results and the actual power load, which is difficult to meet the actual needs of long-term planning and dispatching decision of the power system. SUMMARY

[0003] In order to solve the above problems existing in the prior art, the present application proposes a power load prediction method based on multi-scale frequency domain state space model to solve the problems of low efficiency of periodic feature extraction, serious noise interference and high calculation complexity existing in the prior art power load prediction method, so as to realize high-precision long-term prediction of high-dimensional and non-stationary power load time series by fusing frequency domain analysis and multi-scale modeling architecture.

[0004] The application achieves the above-mentioned purposes by adopting the following technical scheme: The power load prediction method based on a multi-scale frequency domain state space model has the characteristics that the method comprises the following steps: Step 1: collecting power load data and preprocessing the power load data to obtain a preprocessed power load data set , wherein, represents the nth power load data in the preprocessed power load data set, L is the length of the sequence, and N is the feature dimension of the power load; Step 2: constructing a power load prediction network, comprising a preprocessing module, a double-path feature extraction module, a multi-scale prediction module, a trend component learning module, a linear fusion layer, and processing to obtain predicted power load data ; wherein H represents the length of the prediction window; Step 2.1, the preprocessing module comprises a position encoding layer, a normalization processing layer, and a data decomposition layer, and processes to obtain the nth trend term and the nth periodic term ; The position encoding layer encodes to obtain the nth position encoding result ; The normalization processing layer adopts the RevIN method to normalize to obtain the nth normalized power load data ; The data decomposition layer decomposes into the nth trend term and the nth periodic term by a multi-kernel moving average decomposition method; Step 2.2, the double-path feature extraction module utilizes fast Fourier transform to process in length and dimension respectively to obtain the nth periodic term represented by a frequency domain complex value ; Step 2.3, the multi-scale prediction module comprises a coarse-grained modeling unit, a fine-grained modeling unit, and a feature fusion and residual connection unit, and processes the nth periodic term represented by a frequency domain complex value to obtain the periodic term result of the nth power load ; Step 2.4, the trend component learning module comprises a linear layer, and projects the nth trend term to the prediction window length H by using a linear network to obtain the result of the nth trend term ; Step 2.5, the linear fusion layer will and After fusion, inverse normalization is performed to obtain the nth power load data with a prediction window length H. ; Step 3: Train the power load prediction network using stochastic gradient descent and calculate the network loss using MSE loss. Stop training when the loss converges or the maximum number of iterations is reached, thereby generating the power load prediction model corresponding to the optimal parameters, which is used to predict future long-term power load data.

[0005] The power load forecasting method based on a multi-scale frequency domain state-space model described in this invention is characterized in that step 2.3 includes the following steps: Step 2.3.1, coarse-grained modeling units will real part Projecting to a larger dimension The above yields the nth coarse-grained real part periodic feature. Meanwhile, the residual terms are retained. Then, through the positive state-space model, By processing the time dimension, the real part periodicity feature of the nth coarse-grained time dimension is obtained. Finally, the inverse state-space model is used to... By processing the channel dimensions, we obtain the real part periodicity feature of the nth coarse-grained channel dimension. ; Step 2.3.2, fine-grained modeling unit will real part Projecting to a smaller dimension The above yields the nth fine-grained periodic feature of the real part. Meanwhile, the residual terms are retained. Then, through the positive state-space model, By processing the time dimension, the real part periodicity feature of the nth fine-grained time dimension is obtained. Finally, the inverse state-space model is used to... By processing the channel dimensions, we obtain the real part periodicity feature of the nth fine-grained channel dimension. ; Step 2.3.3, Feature Fusion and Residual Connection Unit will and The fusion is performed to obtain the nth coarse-grained fusion real part periodic feature result. ; Will and After fusion, it is then combined with the residual term of the fine-grained layer. After residual connection, projection is made to the coarse granularity, so as to obtain the n th fine-grained fusion real part periodic feature result ; After fusion , residual connection is made with the coarse-grained layer residual term , projection is made to the prediction window length H, so as to obtain the n th real part periodic term prediction result . Step 2.3.4, according to the process of steps 2.3.1-2.3.3, the imaginary part is input into the multi-scale prediction module for processing, so as to obtain the n th imaginary part periodic term prediction result . Step 2.3.5, data frequency domain reconstruction and time domain conversion are made on and , so as to obtain the periodic complex signal of the n th power load . After the fast inverse Fourier transform is made on to convert it back to the time domain, the periodic term result of the n th power load is obtained .

[0006] The electronic device comprises a memory and a processor, wherein the memory is used for storing a program supporting the processor to execute the long-term power load prediction method, and the processor is configured to execute the program stored in the memory.

[0007] The computer readable storage medium stores a computer program, wherein the computer program is executed by the processor to execute the steps of the long-term power load prediction method.

[0008] Compared with the prior art, the present application has the following advantages: 1. The present application uses efficient feature extraction driven by frequency domain analysis, breaks through the noise bottleneck of time domain modeling, and decomposes the power load sequence into the frequency domain through double-path fast Fourier transform, effectively isolates the periodic mode and noise compared with the traditional time domain model, greatly improves the feature extraction efficiency. The mechanism does not need to rely on a complex time domain attention mechanism, directly filters redundant information from the frequency domain, so that the model improves the prediction accuracy in the high-noise power load sequence.

[0009] 2、The application balances global and local feature expression by using hierarchical dependent modeling of the multi-scale Mamba architecture. Unlike the quadratic complexity and single-scale modeling of attention mechanism-based models, the application adopts a quadruple Mamba block and double-scale feature layer to capture coarse-grained global trends (such as weekly / monthly periodicity) and fine-grained local fluctuations (such as hourly-level mutations), thereby capturing the change characteristics of power load data more accurately and improving the prediction efficiency.

[0010] 3、The linear complexity and hardware-aware design of the application balance prediction accuracy and real-time performance while adaptively decomposing and fusing across dimensions to enhance the adaptability of complex data. The application combines the complexity of linear scanning mechanism and frequency domain analysis to reduce the calculation overhead while ensuring the accuracy of the prediction. Compared with traditional single decomposition methods, the application dynamically separates the trend item and periodic component by using multiple kernel moving average, and fuses the frequency domain features by using residual connection, which performs better in non-stationary data. BRIEF DESCRIPTION OF DRAWINGS

[0011] Figure 1 is a flowchart of the training method of the power load prediction model of the application; Figure 2 is a structural diagram of the power load prediction model of the application. DETAILED DESCRIPTION

[0012] In this embodiment, a power load prediction method based on a multi-scale frequency domain state space model, as shown in Figure 1 , comprises the following steps: Step 1: Collect power load data and perform preprocessing to obtain a preprocessed power load data set , wherein represents the nth power load data in the preprocessed power load data set, L is the length of the sequence, and N is the feature dimension of the power load; Step 2: As shown in Figure 2 , a power load prediction network is constructed, including a preprocessing module, a double-path feature extraction module, a multi-scale prediction module, a trend component learning module, and a linear fusion layer, and the power load data is processed to obtain predicted power load data ; wherein H represents the length of the prediction window; in this example, L=96, H=96, 192, 288, 384, and N=7.

[0013] Step 2.1, the preprocessing module includes a position encoding layer, a normalization processing layer, and a data decomposition layer, and the power load data is processed to obtain the nth trend item and the nth periodic item ; The position coding in this example uses a fixed coding mode; Position coding layer pair Position coding is performed to obtain the nth position coding result ; The normalization processing layer uses the RevIN method to perform normalization processing on to obtain the nth normalized power load data ; The data decomposition layer decomposes into the nth trend term and the nth periodic term ; Step 2.2, as shown in Figure 2 , the double-path feature extraction module uses the fast Fourier transform to process in length and dimension respectively, to obtain the nth periodic term represented by a frequency domain complex value ; Step 2.3, the multi-scale prediction module includes a coarse-grained modeling unit, a fine-grained modeling unit, a feature fusion and residual connection unit, and processes the nth periodic term represented by a frequency domain complex value to obtain the nth periodic term result of the power load Step 2.3.1, the coarse-grained modeling unit projects the real part of to a larger dimension to obtain the nth coarse-grained real part periodic feature , while retaining the residual term , then processes in the time dimension through the forward state space model to obtain the nth coarse-grained time-dimension real part periodic feature , and finally processes in the channel dimension through the backward state space model to obtain the nth coarse-grained channel-dimension real part periodic feature ; Step 2.3.2, the fine-grained modeling unit projects the real part of to a smaller dimension to obtain the nth fine-grained real part periodic feature , while retaining the residual term , then processes in the time dimension through the forward state space model to obtain the nth fine-grained time-dimension real part periodic feature , and finally processes The channel dimension is processed to obtain a real part periodicity feature of the nth fine-grained channel dimension ; The feature fusion and residual connection unit of step 2.3.3 fuses and to obtain an nth coarse-grained fused real part periodicity feature result ; After fusing and , the fine-grained layer residual term is connected in residual, and then projected to the coarse-grained layer to obtain an nth fine-grained fused real part periodicity feature result ; After fusing and , the coarse-grained layer residual term is connected in residual, and then projected to the prediction window length H to obtain an nth real part periodicity term prediction result ; Step 2.3.4, according to the processes of steps 2.3.1-2.3.3, the imaginary part is input into the multi-scale prediction module for processing to obtain an nth imaginary part periodicity term prediction result ; Step 2.3.5, data frequency domain reconstruction and time domain conversion are performed on and to obtain a periodicity complex signal of the nth power load , and after converting back to the time domain through the fast inverse Fourier transform, an nth power load periodicity term result is obtained ; Step 2.4, the trend component learning module includes a linear layer, which projects the nth trend term to the prediction window length H using a linear network to obtain an nth trend term result ; Step 2.5, a linear fusion layer fuses and after which de-normalization is performed to obtain an nth power load data of the prediction window length H ; Step 3, the random gradient descent method is used to train the power load prediction network, and the MSE loss is used to calculate the network loss until the loss converges or the maximum number of iterations is reached, and the training is stopped, thereby generating a power load prediction model corresponding to the optimal parameters for predicting future long-term power load data.

[0014] In this embodiment, an electronic device includes a memory for storing a program supporting a processor to execute the above method, and the processor configured to execute the program stored in the memory.

[0015] In this embodiment, a computer readable storage medium has a computer program stored thereon, and the computer program is run by a processor to perform the steps of the above method.

Claims

1. A power load forecasting method based on a multi-scale frequency domain state-space model, characterized in that, Includes the following steps: Step 1: Collect and preprocess power load data to obtain a preprocessed power load dataset. ,in, This represents the nth power load data in the preprocessed power load dataset, where L is the length of the sequence and N is the feature dimension of the power load. Step 2: Construct a power load forecasting network, including: a preprocessing module, a dual-path feature extraction module, a multi-scale forecasting module, a trend component learning module, and a linear fusion layer, and then... The data is processed to obtain the predicted power load data. Where H represents the length of the prediction window; Step 2.1: The preprocessing module includes: a position encoding layer, a normalization processing layer, and a data decomposition layer, and performs... After processing, the nth trend term is obtained. and the nth periodic term ; The location encoding layer Perform position encoding to obtain the nth position encoding result. ; The normalization processing layer uses the RevIN method to process... Normalization is performed to obtain the nth normalized power load data. ; The data decomposition layer uses a multi-kernel moving average decomposition method to... Decomposed into the nth trend term and the nth periodic term ; Step 2.2: The dual-path feature extraction module uses Fast Fourier Transform to extract features in both length and dimension. After processing, the periodic term in the nth frequency domain complex-valued representation is obtained. ; Step 2.3: The multi-scale prediction module includes: a coarse-grained modeling unit, a fine-grained modeling unit, and a feature fusion and residual connection unit, and it also handles the periodic terms of the nth frequency domain complex-valued representation. The process is performed to obtain the periodic term result of the nth power load. ; Step 2.4: The trend component learning module includes a linear layer, which processes the nth trend item... By projecting a linear network onto the prediction window length H, the result of the nth trend term is obtained. ; Step 2.5, the linear fusion layer will and After fusion, inverse normalization is performed to obtain the nth power load data with a prediction window length H. ; Step 3: Train the power load prediction network using stochastic gradient descent and calculate the network loss using MSE loss. Stop training when the loss converges or the maximum number of iterations is reached, thereby generating the power load prediction model corresponding to the optimal parameters, which is used to predict future long-term power load data.

2. The power load forecasting method based on a multi-scale frequency domain state-space model according to claim 1, characterized in that, Step 2.3 includes the following steps: Step 2.3.1, coarse-grained modeling units will real part Projecting to a larger dimension The above yields the nth coarse-grained real part periodic feature. Meanwhile, the residual terms are retained. Then, through the positive state-space model, By processing the time dimension, the real part periodicity feature of the nth coarse-grained time dimension is obtained. Finally, the inverse state-space model is used to... By processing the channel dimensions, we obtain the real part periodicity feature of the nth coarse-grained channel dimension. ; Step 2.3.2, fine-grained modeling unit will real part Projecting to a smaller dimension The above yields the nth fine-grained periodic feature of the real part. Meanwhile, the residual terms are retained. Then, through the positive state-space model, By processing the time dimension, the real part periodicity feature of the nth fine-grained time dimension is obtained. Finally, the inverse state-space model is used to... By processing the channel dimensions, we obtain the real part periodicity feature of the nth fine-grained channel dimension. ; Step 2.3.3, Feature Fusion and Residual Connection Unit will and The fusion is performed to obtain the nth coarse-grained fusion real part periodic feature result. ; Will and After fusion, it is then combined with the residual term of the fine-grained layer. After performing residual connections, the result is projected onto the coarse-grained model to obtain the nth fine-grained fused real-part periodic feature result. ; Will and After fusion, it is then combined with the residual term of the coarse-grained layer. After performing residual connection, the result is projected onto the prediction window length H to obtain the prediction result of the nth real part periodic term. ; Step 2.3.4: Following the process of steps 2.3.1-2.3.3, convert the imaginary part... The input is processed in the multi-scale prediction module to obtain the prediction result of the nth imaginary periodic term. ; Step 2.3.5, will and By performing frequency domain reconstruction and time domain transformation on the data, the periodic complex signal of the nth power load is obtained. Then, through fast inverse Fourier transform, After switching back to the time domain, the periodic term result of the nth power load is obtained. .

3. An electronic device, comprising a memory and a processor, characterized in that, The memory is used to store programs that support the processor in executing the long-term power load forecasting method of claim 1 or 2, and the processor is configured to execute the programs stored in the memory.

4. A computer-readable storage medium storing a computer program thereon, characterized in that, The computer program, when run by a processor, performs the steps of the long-term power load forecasting method as described in claim 1 or 2.

Images