Short-term load prediction method and device and medium

Abstract

The invention discloses a short-term load prediction method and device based on an ICEEMDAN-PE model and an IDBO-Informer model, and a medium. According to the method, a combined prediction model based on the ICEEMDAN-PE model and the IDBO-Informer model is provided. According to the model, firstly, the WSTD technology is used for preprocessing load data, and influences of noise and abnormal values are eliminated; then, carrying out multi-scale decomposition on the denoised load data by adopting an ICEEMDAN method, and extracting feature components of different time scales; the complexity of each component is evaluated by using a PE algorithm, and a basis is provided for selection and parameter setting of a subsequent prediction model; and finally, parameters of the Informer prediction model are optimized by adopting IDBO, and the prediction efficiency and accuracy are improved through innovative designs such as a ProbSparse self-attention mechanism, a self-attention distillation mechanism and a generative decoder.

Description

Technical Field

[0001] The present application relates to the technical field of power system load forecasting, and in particular to a short-term load forecasting method, device and medium. Background Art

[0002] In the power system, load forecasting is a key link to ensure the safe and stable operation of the power grid, optimize resource allocation, and improve economic benefits. As an important indicator reflecting the supply and demand relationship of the power system, the accurate prediction of power load is of vital importance to the production scheduling, energy management, and market strategy formulation of power companies. With the development of computer technology and artificial intelligence, machine learning algorithms have gradually been applied to the field of load forecasting, such as support vector machine (SVM), random forest (RF), gradient boosting tree (GBDT), etc. These algorithms have achieved significant improvement in prediction accuracy by extracting nonlinear relationships in load data. However, these algorithms still have limitations in feature selection and model generalization, and it is difficult to fully mine the deep information and complex patterns in load data. The application of deep learning models in load forecasting also faces many challenges. First, load data usually contains a lot of noise and outliers. These interference information will affect the training effect and prediction accuracy of the model; second, load data is highly nonlinear and non-stationary, requiring the model to have strong feature extraction and pattern recognition capabilities; finally, the training process of deep learning models is complex and computationally intensive, requiring high hardware resources, and the optimization and adjustment of model parameters is also a time-consuming and difficult process. Summary of the invention

[0003] The present application provides a short-term load forecasting method, device and medium based on ICEEMDAN-PE and IDBO-Informer models, which has the advantage of effectively improving the accuracy and stability of short-term load forecasting, realizing accurate forecasting of short-term loads, and providing strong support for the safe and stable operation of the power system.

[0004] The technical solution of this application is as follows:

[0005] On the one hand, the present application provides a short-term load forecasting method based on ICEEMDAN-PE and IDBO-Informer models, comprising the following steps:

[0006] S1: Use wavelet soft threshold denoising (WSTD) technology to remove noise from the original load data and improve data quality;

[0007] S2: The denoised data is decomposed at multiple scales using the improved complete ensemble empirical mode decomposition (ICEEMDAN) method to obtain multiple intrinsic mode function (IMF) components and residual terms to capture the characteristics of different time scales.

[0008] S3: The permutation entropy PE algorithm is used to evaluate the complexity of each IMF component;

[0009] S4: Improve the Informer model and use the improved dung beetle optimization IDBO algorithm to optimize the parameters of the Informer model, including integrating chaos and reverse learning strategies for population initialization, introducing adaptive step size and convex lens inverse imaging strategy and random difference mutation strategy to enhance global search capabilities and avoid local optimality;

[0010] S5: Input the decomposed data into the optimized Informer model for prediction, and integrate the prediction results of each IMF component according to the weighted PE value to obtain the final load prediction value.

[0011] Furthermore, in step S1, WSTD is used for data preprocessing; the wavelet can be scaled and time-shifted by the scale factor a and the time-shift factor b to generate a series of sub-wavelets to extract target features at different resolutions; the specific steps of WSTD are as follows:

[0012] The load sequence is denoised and modeled. It is assumed that the pure signal of the load sequence is f(t), and the sample signal of the load sequence containing noise is x(t):

[0013] x(t)=f(t)+σn(t)(1)

[0014] Where t is the time point; n(t) is the probability distribution that follows the normal distribution N(0,σ 2 )’s Gaussian white noise;

[0015] In order to obtain the wavelet sequence equation using the dyadic wavelet, a and b are substituted into the wavelet sequence obtained by scaling and time shifting the wavelet:

[0016]

[0017] In the formula, ψ m,n (t) is the wavelet sequence; ψ(t) represents the wavelet function;

[0018] In order to decompose the sample signal using DWT, a and b are incorporated into the CWT definition of wavelet ψ to obtain the corresponding DWT;

[0019]

[0020] Where W x Represents wavelet transform coefficients; Mallat algorithm is often used in DWT, and scaling functions and wavelet equations are often used as low-pass and high-pass filters; Mallat algorithm is used to perform multi-level decomposition of DWT; after decomposition, the corresponding scaling coefficients and wavelet coefficients are obtained; scaling coefficients contain low-frequency information, and wavelet coefficients contain high-frequency information:

[0021]

[0022] In the formula, c m,n is the scale factor; d m,n are wavelet coefficients;

[0023] The threshold λ is calculated to distinguish the noise signal from the valid signal. According to the different threshold functions, WTD is divided into wavelet hard threshold denoising in equation (5) and WSTD in equation (6):

[0024]

[0025] Where W x is the wavelet transform coefficient; λ is the threshold; |W x | is the evaluated wavelet transform coefficient; the noise wavelet transform coefficient is set to zero and the load time series is reconstructed.

[0026] Furthermore, in step S2:

[0027] CEEMDAN introduces adaptive noise and runs EMD multiple times to obtain multiple groups of modal functions, which are then averaged to obtain a more accurate decomposition result. The formula is as follows:

[0028]

[0029] In the formula, ω (i) is Gaussian white noise; r k is the kth group of residuals; I is the generated ω (i) ; M(·) is the local mean of the generated sequence.

[0030] Furthermore, in step S3:

[0031] The PE algorithm is used to measure the complexity of each sub-module component;

[0032] The steps of PE algorithm are as follows:

[0033] (1) The reconstructed sequence of the phase space of the time series can be obtained by the following formula:

[0034] Y s =[x(s),x(s+τ),…x(s+(m-1)τ)](9)

[0035] Where s = 1, 2, ..., N-(m-1)τ, τ is the delay time, and m is the embedding dimension;

[0036] (2) For each reconstructed vector Y s Rearrange the components of from small to large to get the symbol sequence:

[0037] Y(g)=(s1,s2,…,s m ), g=1,2,…,landl≤m! (10)

[0038] (3) The probability of each symbol sequence Y(g) occurring is calculated as P g :

[0039]

[0040] Where N(Y(g)) is the number of Y(g), P g satisfy

[0041] (4) The PE of this time series is defined as:

[0042]

[0043] As can be seen from formula (12), PE is an indicator to measure the complexity of a sequence. The larger the PE value, the more complex the sequence. The complexity of each subsequence is evaluated by calculating the PE value of each IMF.

[0044] Further, in step S4:

[0045] The improvements of Informer include: using probabilistic sparse self-attention mechanism to reduce the complexity of matrix calculation; introducing self-attention distillation mechanism to extract the main features of time series; using decoder to directly generate output prediction value; the formula of ProbSparse self-attention is:

[0046]

[0047] Where: is a sparse matrix of the same size as Q;

[0048] The self-attention distillation operation is introduced, and the convolution and pooling operations are used to reduce the dimension and network parameters; the distillation process from the jth layer to the j+1th layer is as follows:

[0049]

[0050] In the formula, [·] AB represents the attention block, which contains multi-head ProbSparse self-attention and basic operations; Conv1d(·) uses the ELU activation function to perform one-dimensional convolution filtering on the time dimension; represents downsampling;

[0051] The input of the decoder is represented as:

[0052]

[0053] In the formula, Start token; Placeholder for the target sequence that sets the scalar to 0; L token is the length of the input; L y is the length of the output sequence; d model is the dimension of the model input; in the ProbSparse self-attention calculation, masked multi-head attention is applied by setting the masked dot product to -∞.

[0054] Further, in step S4:

[0055] The IDBO algorithm is an improvement on the DBO algorithm. The IDBO algorithm combines the chaos initialization method with the reverse learning initialization strategy to improve the convergence speed of DBO. The specific steps are as follows:

[0056] Fuch mapping is used to generate the initial population of DBO.

[0057]

[0058] Where P(t) is the chaotic variable of Fuch mapping. When the Fuch chaotic mapping is used to generate the initial solution, the reverse learning strategy is introduced. The mathematical expression of the reverse learning strategy is shown in formula (17):

[0059] OP i =K(P min +P max )-P i (17)

[0060] In the formula, OP i is the inverse solution P corresponding to each initial solution i ;P min and P max Respectively represent the minimum and maximum values ​​in the initial solution; K is a random value in the range of (0,1);

[0061] A dynamic selection strategy is adopted to update the target position alternately with the adaptive step size strategy and the convex lens imaging inversion strategy under a certain probability; the strategy is mainly determined by the linearly decreasing adaptive step control factor α0, as follows:

[0062]

[0063] At the same time, the convex lens imaging learning strategy is introduced to disturb the foraging dung beetle population. The equation is:

[0064]

[0065] Among them, P j Represents the component of the current individual in dimension j; a j and bj They represent the minimum and maximum values ​​of the j dimension in the search space respectively; Expressing the inverse solution of the convex lens P j ;

[0066] Then, the choice of which strategy to use for the target position update is determined by the selection probability P s Determine, as shown in formula (20); if the random value P s <0.5, the random stepping strategy is used to update the position of the foraging dung beetle; otherwise, the convex lens reverse learning strategy is used to update the position, as shown in formula (21):

[0067]

[0068] A random difference mutation strategy was introduced to enhance the population diversity of dung beetles; the equation is as follows:

[0069] P(t+1)=r1×(P * -P(t))+r2×(P′(t)-P(t)) (22)

[0070] Where P(t+1) is a new individual obtained by random differential mutation, r1 and r2 are random values ​​in the range of [0, 1]; * is the current optimal individual position; P′(t) is the position of a randomly selected individual in the population.

[0071] On the other hand, the present application provides a short-term load forecasting device based on ICEEMDAN-PE and IDBO-Informer models, including a processor and a memory, wherein the memory stores a computer program, and when the computer program is called and executed by the processor, the method as described above is implemented.

[0072] On the other hand, the present application provides a computer-readable medium, wherein the computer-readable medium stores a computer program, and when the computer program is called and executed by a computer, the method as described above is implemented.

[0073] In summary, the beneficial effects of the present application are: through wavelet soft threshold denoising, ICEEMDAN decomposition, permutation entropy complexity evaluation and IDBO algorithm optimization, combined with the Informer prediction model, the combined model proposed in the present invention shows excellent performance in short-term load forecasting. The model is superior to traditional machine learning models and other deep learning models in terms of prediction accuracy, smoothness and trend consistency. The present invention introduces innovative algorithms such as ICEEMDAN and IDBO, which effectively solves the problems of mode mixing and local optimality in traditional methods. The successful application of these technologies not only enriches the theoretical system of short-term load forecasting, but also provides strong support for practical applications. By integrating a variety of advanced technologies and algorithms, the model effectively improves the accuracy and stability of short-term load forecasting, realizes accurate prediction of short-term loads, and provides strong support for the safe and stable operation of power systems. BRIEF DESCRIPTION OF THE DRAWINGS

[0074] Figure 1 It is a schematic diagram of the WSTD-ICEEMDAN-PE-IDBO-NLSTM model;

[0075] Figure 2 This is the DWT schematic diagram;

[0076] Figure 3 It is the overall framework of Informer;

[0077] Figure 4 is the test function convergence curve;

[0078] Figure 5 is the original load data;

[0079] Figure 6 It is the comparison of load curves before and after noise reduction;

[0080] Figure 7 It is the comparison of WSTD time-frequency diagram before and after noise reduction;

[0081] Figure 8 is the subsequence decomposed by ICEEMDAN;

[0082] Fig. 9 It is the result of load decomposition and then PE reconstruction;

[0083] Fig.10 are the prediction curves of different models;

[0084] Fig.11 are the error curves of different models;

[0085] Fig.12 are the prediction curves of different optimization algorithms;

[0086] Fig.13is the absolute error of different optimization algorithms;

[0087] Fig.14 It is an evaluation index for different modal decomposition methods. DETAILED DESCRIPTION

[0088] The specific implementation of the present application is described in detail below with reference to the accompanying drawings.

[0089] A specific implementation of the present application provides a short-term load forecasting method based on ICEEMDAN-PE and IDBO-Informer models, referring to Figure 1 , including the following steps:

[0090] S1: Use wavelet soft threshold denoising (WSTD) technology to remove noise from the original load data and improve data quality;

[0091] S2: The denoised data is decomposed at multiple scales using the improved complete ensemble empirical mode decomposition (ICEEMDAN) method to obtain multiple intrinsic mode function (IMF) components and residual terms to capture the characteristics of different time scales.

[0092] S3: The permutation entropy PE algorithm is used to evaluate the complexity of each IMF component;

[0093] S4: Improve the Informer model and use the improved dung beetle optimization IDBO algorithm to optimize the parameters of the Informer model, including integrating chaos and reverse learning strategies for population initialization, introducing adaptive step size and convex lens inverse imaging strategy and random difference mutation strategy to enhance global search capabilities and avoid local optimality;

[0094] S5: Input the decomposed data into the optimized Informer model for prediction, and integrate the prediction results of each IMF component according to the weighted PE value to obtain the final load prediction value.

[0095] According to steps S1-S5, a combined prediction model based on ICEEMDAN-PE and IDBO-Informer is constructed.

[0096] In step S1, WSTD is used for data preprocessing, which can effectively eliminate noise in the data while retaining useful information and avoiding information loss. Wavelets can be scaled and time-shifted by scale factor a and time-shift factor b to generate a series of sub-wavelets to extract target features at different resolutions; the specific steps of WSTD are as follows:

[0097] In order to perform denoising modeling on the load sequence, it is assumed that the pure signal of the load sequence is f(t), and the sample signal of the load sequence containing noise is x(t):

[0098] x(t)=f(t)+σn(t) (1)

[0099] Where t is the time point; n(t) is the probability distribution that follows the normal distribution N(0,σ 2 ) is a Gaussian white noise.

[0100] In order to obtain the wavelet sequence equation using the dyadic wavelet, a and b are substituted into the wavelet sequence obtained by scaling and time shifting the wavelet:

[0101]

[0102] In the formula, ψ m,n (t) is the wavelet sequence; ψ(t) represents the wavelet function.

[0103] In order to decompose the sample signal using DWT, a and b are incorporated into the CWT definition of wavelet ψ to obtain the corresponding DWT;

[0104]

[0105] Where W x Represents wavelet transform coefficients.

[0106] Mallat algorithm is often used in DWT, and scaling function and wavelet equation are often used as low-pass and high-pass filters; Mallat algorithm is used to perform multi-level decomposition of DWT, the process is as follows Figure 2 As shown; after decomposition, the corresponding scale coefficient and wavelet coefficient are obtained. The scale coefficient contains low-frequency information, and the wavelet coefficient contains high-frequency information:

[0107]

[0108] In the formula, c m,n is the scale factor; d m,n are the wavelet coefficients.

[0109] Calculate the threshold λ to distinguish the noise signal from the valid signal; in wavelet soft threshold denoising, the pre-setting of the threshold is essential, which will determine the effect of the denoising process. According to the different threshold functions, WTD is divided into wavelet hard threshold denoising in formula (5) and WSTD in formula (6):

[0110]

[0111]

[0112] Where W x is the wavelet transform coefficient; λ is the threshold; |W x | is the evaluated wavelet transform coefficient; the noise wavelet transform coefficient is set to zero and the load time series is reconstructed.

[0113] In step S2, fully integrated empirical mode decomposition adaptive noise (CEEMDAN) is an improved version of EMD for signal processing and data decomposition. EMD is an adaptive data decomposition method that decomposes the signal into a series of IMF components and residual terms. However, the traditional EMD method may have a mode mixing problem, resulting in inaccurate decomposition results. To solve this problem, in this embodiment, CEEMDAN introduces adaptive noise, and runs EMD multiple times to obtain multiple groups of modal functions, which are then averaged to obtain more accurate decomposition results. Therefore, in order to further enhance CEEMDAN and solve its limitations, the ICEEMDAN method is proposed. ICEEMDAN introduces adaptive noise by optimization, solves the "pseudo" mode problem, and achieves more accurate and stable data decomposition. Based on multiple runs of EMD, ICEEMDAN effectively solves some limitations of CEEMDAN by improving noise strategies and related processing. This improves the accuracy and reliability of signal decomposition.

[0114] The formula of ICEEMDAN method is as follows:

[0115]

[0116] In the formula, ω (i) is Gaussian white noise; r k is the kth group of residuals; I is the generated ω (i) ; M(·) is the local mean of the generated sequence.

[0117] In step S3, the PE algorithm is a method for detecting the randomness of a time series by effectively amplifying the slight changes in the time series. The algorithm is simple, has strong anti-interference ability, and good robustness. The PE algorithm is used to measure the complexity of each sub-mode component.

[0118] The steps of PE algorithm are as follows:

[0119] (1) The reconstructed sequence of the phase space of the time series can be obtained by the following formula:

[0120] Y s =[x(s),x(s+τ),…x(s+(m-1)τ)](9)

[0121] Where s = 1, 2, ..., N-(m-1)τ, τ is the delay time, and m is the embedding dimension;

[0122] (2) For each reconstructed vector Y s Rearrange the components of from small to large to get the symbol sequence:

[0123] Y(g)=(s1,s2,…,s m ), g=1,2,…,landl≤m! (10)

[0124] (3) The probability of each symbol sequence Y(g) occurring is calculated as P g :

[0125]

[0126] Where N(Y(g)) is the number of Y(g), P g satisfy

[0127] (4) The PE of this time series is defined as:

[0128]

[0129] As can be seen from formula (12), PE is an indicator to measure the complexity of a sequence. The larger the PE value, the more complex the sequence. Therefore, the complexity of each subsequence is evaluated by calculating the PE value of each IMF.

[0130] The Informer network is a variant of the Transformer, which effectively solves the problem of long sequence prediction. In step S4, the improvements of the Informer include: using a probabilistic sparse self-attention mechanism to reduce the complexity of matrix calculation; introducing a self-attention distillation mechanism to extract the main features of the time series, effectively reducing memory usage; using the decoder to directly generate output prediction values ​​to achieve the purpose of long sequence prediction. The structure of the Informer model is as follows: Figure 3 As shown. The formula of ProbSparse self-attention can be derived as:

[0131]

[0132] Where: is a sparse matrix of the same size as Q; this allows ProbSparse self-attention to only compute dot products for each query-key pair. This approach effectively solves the quadratic computational complexity problem of the Transformer self-attention mechanism.

[0133] The self-attention distillation operation is introduced, and the convolution and pooling operations are used to reduce the dimensionality and network parameters, thereby greatly reducing the overall spatial complexity. This overcomes the obstacle of not being able to stack too long inputs. The convolution and pooling operations are used to reduce the dimensionality and network parameters; the distillation process from the jth layer to the j+1th layer is as follows:

[0134]

[0135] In the formula, [·] ABrepresents the attention block, which contains multi-head ProbSparse self-attention and basic operations; Conv1d(·) uses the ELU activation function to perform one-dimensional convolution filtering on the time dimension; Represents downsampling; this includes multi-head probabilistic sparse self-attention as well as key operations in the attention block. This will solve the problem of high memory usage in the Transformer model.

[0136] This embodiment further proposes a generative decoder. This gets all the predictions in one step, rather than the traditional step-by-step approach. This speeds up predictions for long sequences because only one forward step is needed to generate the entire output sequence. It also avoids error propagation in the prediction phase.

[0137] The input of the decoder is represented as:

[0138]

[0139] In the formula, Start token; Placeholder for the target sequence that sets the scalar to 0; L token is the length of the input; L y is the length of the output sequence; d model is the dimension of the model input; in the ProbSparse self-attention calculation, masked multi-head attention is applied by setting the masked dot product to -∞. It prevents each position from participating in the next position, thus avoiding auto-regression. This solves the efficiency problem of predicting long-term outputs in the transformer.

[0140] In step S4, the IDBO algorithm is an improvement on the DBO algorithm. Although the DBO algorithm has good performance and has been effectively applied to some engineering design problems, it also has disadvantages such as low global searchability and premature convergence to local optimum. This embodiment proposes an enhanced DBO algorithm with a specific enhancement strategy to solve these problems. The population initialization in DBO is randomly generated, which will lead to uneven distribution of population initialization and diversity of the initialized population. The chaos and reverse learning strategies are integrated to initialize the population. In this embodiment, the IDBO algorithm combines the chaos initialization method with the reverse learning initialization strategy to improve the convergence speed of DBO.

[0141] Common chaotic maps include Tent map and Logistic map. However, both are finite folding chaotic maps with poor ergodicity and uniformity. Fuch is an infinite foldable chaotic map, which has better ergodicity, dynamics and convergence than traditional chaotic maps.

[0142] Fuch mapping is used to generate the initial population of DBO.

[0143]

[0144] In the formula, P(t) is the chaotic variable of Fuch mapping. When the Fuch chaotic mapping is used to generate the initial solution, the reverse learning strategy is introduced to expand the search space of the dung beetle and improve the quality of the initial solution of the dung beetle. The mathematical expression of the reverse learning strategy is shown in formula (17):

[0145] OP i =K(P min +P max )-P i (17)

[0146] In the formula, OP i is the inverse solution P corresponding to each initial solution i ;P min and P max They represent the minimum and maximum values ​​in the initial solution respectively; K is a random value in the range of (0,1).

[0147] Since DBO adopts a random strategy, it lacks adaptive ability during the foraging stage of dung beetles, has weak global search ability, and is prone to falling into local optimality. In order to further improve the optimization performance of the algorithm, a dynamic selection strategy is adopted, and the adaptive step size strategy and the convex lens imaging inversion strategy are used to update the target position alternately under a certain probability. In the initial iteration, a larger step size can expand the foraging search range, improve the global search ability of the algorithm, and help the algorithm find a better solution faster and accelerate convergence. In the later iterations, a smaller step size is conducive to the local search of the algorithm. The regular change of the step size plays a guiding role in finding foraging dung beetles, which means that in the overall context, there is a gradual transition from global search to local search. The strategy is mainly determined by the linearly decreasing adaptive step control factor α0, as follows:

[0148]

[0149] At the same time, the convex lens imaging learning strategy is introduced to disturb the foraging dung beetle population. The equation is:

[0150]

[0151] Among them, P j Represents the component of the current individual in dimension j; a j and b j They represent the minimum and maximum values ​​of the j dimension in the search space respectively; Expressing the inverse solution of the convex lens P j .

[0152] Then, the choice of which strategy to use for the target position update is determined by the selection probability P sDetermine, as shown in formula (20); if the random value P s <0.5, the random stepping strategy is used to update the position of the foraging dung beetle; otherwise, the convex lens reverse learning strategy is used to update the position, as shown in formula (21):

[0153]

[0154] The position update equation of the stealing dung beetle updates its position based on the individual's current optimal state, which may lead to a reduction in population diversity. This causes the algorithm to fall into a local optimum and the convergence accuracy is not high. Therefore, a random difference mutation strategy is introduced to enhance the population diversity of dung beetles; the equation is as follows:

[0155] P(t+1)=r1×(P * -P(t))+r2×(P′(t)-P(t)) (22)

[0156] Where P(t+1) is a new individual obtained by random differential mutation, r1 and r2 are random values ​​in the range of [0, 1]; * is the current optimal individual position; P′(t) is the position of a randomly selected individual in the population.

[0157] In order to verify the advantages of the method proposed in the present invention, the IDBO algorithm is comprehensively compared with the particle swarm optimization algorithm (PSO), sparrow optimization (SSA) algorithm, gray wolf optimization (GWO) algorithm, and the original DBO algorithm. In order to avoid the randomness of the experimental results, the population size of the PSO, SSA, GWO, DBO, and IDBO algorithms is set to 30, and the maximum number of iterations of each algorithm is 100. The specific parameters of each optimization algorithm are shown in Table 1. The present invention selects 6 standard test functions to test the IDBO algorithm of the present invention. Among them, F1 and F4 are single-peak test functions used to test the speed and accuracy of the global optimal solution of the optimization algorithm. F5, F6, F7, F8, F9, F10, F11, F12, F13, F14, F15, F16, F17, F18, F19, F20, F21, F22, F23, F24, F25, F26, F27, F28, F29, F30, F31, F32, F33, F34, F35, F36, F40, F37, F41, F42, F43, F5 11 、F 15 、F 20 is a multi-peak test function used to test the performance of the optimization algorithm in terms of local search capability. In order to test the search performance of the algorithm, the best value (BV), worst value (WV), mean value (MV) and standard deviation (SD) of the search results are recorded. The experimental results are shown in Figure 4 As shown in Table 2. Figure 4As can be seen from Table 2, the original DBO algorithm cannot solve the complex standard test function well. In contrast, except for F15, the BV, WV, MV, and SD of the IDBO algorithm on the test function are all zero, that is, the global optimum is stably solved. For the multi-peak function F15 with a valley shape, the global optimal point is at the bottom of the valley-finding difficulty, and the DBO algorithm has a good valley-finding accuracy for this function. Therefore, the IDBO algorithm cannot improve the optimization accuracy of this function. Although the F15 function did not find the global optimum, it is still obvious that the three improvement strategies have different degrees of improvement compared with the DBO algorithm. The convergence curves of the five metaheuristic algorithms were drawn based on the experimental data, which can more intuitively observe the effectiveness of the IDBO algorithm, reflecting the convergence speed, stability and ability of the algorithm to get rid of the local optimum.

[0158] Table 1 Parameter settings of each optimization algorithm

[0159]

[0160]

[0161] Table 2 Test results of each optimization algorithm

[0162]

[0163] This embodiment is based on the analysis of the original power load data set in a specific area of ​​Wuxi City, which covers the power consumption from January 1 to January 15, 2020, with a collection interval of 15 minutes, totaling 2400 data points. In order to reduce the complexity of model operation, the data has been normalized. All experiments in this embodiment are carried out in Python 3.8 environment, and the development of deep learning models depends on PyTorch and TensorFlow libraries. The hardware configuration used in the experiment includes: Intel i9-14900KF CPU, NVIDIA 3090GPU and 32GB memory.

[0164] In order to comprehensively evaluate the performance of the proposed short-term load forecasting model based on WSTD-ICEEMDAN-PE-IDBO-Informer, the present invention selects multiple comparative models for empirical analysis. These comparative models include but are not limited to versions based on support vector regression (SVR), long short-term memory network (LSTM), gated recurrent unit (GRU), Transformer model, and different optimization algorithms (such as particle swarm optimization PSO, gray wolf optimization GWO, dung beetle optimization DBO, etc.) combined with the Informer model.

[0165] In the data preprocessing stage, in order to improve the accuracy of load forecasting, the wavelet soft threshold denoising (WSTD) technology is used to reduce the noise of the original load data. The load curve after noise reduction is smoother than the original data, which can more clearly reflect the real trend of load changes and provide a more reliable data basis for subsequent data decomposition and prediction. Figure 5 shows the raw load data, while Figure 6 and Figure 7 The comparison of load curves before and after noise reduction and the comparison of WSTD time-frequency diagrams are respectively displayed, which intuitively shows the noise reduction effect.

[0166] Based on the ICEEMDAN decomposition technology, the power load sequence is finely divided into multiple intrinsic mode function (IMF) components with high to low frequencies, including six IMF components and one residual component (res), such as Figure 8 As shown, the decomposed sequence structure is intuitively displayed. Fig. 9 It further reveals the integration of subsequences after load data decomposition. Among them, the first high-frequency subsequence Sub1 integrates IMF1 and IMF2, the second subsequence Sub2 integrates IMF3 and IMF4, and the subsequent Sub3 and Sub4 correspond to IMF5 and IMF6 respectively. This reorganization method helps to more clearly understand the contribution of different frequency components to the overall load change.

[0167] By comparing the ICEEMDAN-PE-IDBO combined with SVR, LSTM, GRU, Transformer and the ICEEMDAN-PE-IDBO-Informer model proposed in the present invention, it can be seen that the model of the present invention has significant improvements in RMSE, MAE and R2. The comparative experimental evaluation of different models is shown in Table 3. Specifically, compared with the ICEEMDAN-PE-IDBO-SVR model, the RMSE dropped significantly from 296.8 to 109.2, the MAE decreased from 265.7 to 81.3, and the R2 increased from 0.947 to 0.991, showing the great advantages of the model of the present invention in prediction error and fit. Compared with the ICEEMDAN-PE-IDBO-LSTM and ICEEMDAN-PE-IDBO-GRU models, the RMSE and MAE values ​​of the model of the present invention were reduced by about 30% and 20% respectively, and the R2 was increased by nearly 3%, which proves the excellent performance of the model of the present invention in processing complex load data.

[0168] Table 3 Evaluation indicators of different model comparison experiments

[0169]

[0170]

[0171] From the perspective of the smoothness and trend consistency of the prediction curve, the ICEEMDAN-PE-IDBO-Informer model also performs well. Fig.10 As shown in the figure, the prediction curve of the model is most consistent with the actual load curve. The prediction of both peak and trough is relatively accurate, and the overall trend is consistent without obvious deviation or fluctuation. This reflects the strong ability of the model in capturing the law and trend of load changes. Fig.11 The error curve of further confirms the significant advantage of the model of the present invention in error control, and the error fluctuation range is much smaller than that of other models. Fig.12 It can be seen from the prediction curve that the prediction curve of the IDBO-Informer model is most consistent with the actual load data, especially the prediction at the peak and trough is more accurate. Fig.13 The absolute error graph also intuitively demonstrates the significant effect of the IDBO algorithm in reducing prediction errors, and its error fluctuation range is significantly smaller than that of other algorithms.

[0172] In the comparative analysis of optimization algorithms, the performance of six different optimization algorithms combined with the Informer model in short-term load forecasting is compared. The six algorithms are particle swarm optimization (PSO), genetic simulated annealing (GSA), gray wolf optimization (GWO), cuckoo search (CS), dung beetle optimization (DBO) and the improved dung beetle optimization (IDBO) proposed in this paper. Through the comprehensive analysis of the three evaluation indicators of RMSE, MAE and R2, as shown in Table 4, it can be seen that the significant advantages of the IDBO algorithm in optimizing the parameters of the Informer model can be clearly seen. Compared with other optimization algorithms, when IDBO optimizes the Informer model, the RMSE is significantly reduced from 169.9 of PSO to 109.2, a decrease of about 36%; MAE is greatly reduced from 138.7 of PSO to 81.3, a decrease of about 41%; and R2 is increased from 0.974 of PSO to 0.991, an increase of nearly 2%. This shows that the IDBO algorithm performs well in reducing prediction errors and improving model fit. Compared with GSA and GWO algorithms, IDBO also shows significant performance improvement. The RMSE and MAE of IDBO are reduced by about 36% and 37% respectively compared with GSA, and R2 is increased by nearly 1.6%; compared with GWO, the RMSE and MAE are reduced by about 32% and 39% respectively, and R2 is increased by nearly 1.2%. These improvements all show the advantages of IDBO in global search capability and local optimization accuracy. Furthermore, compared with the unimproved dung beetle optimization algorithm (DBO), although DBO has shown good prediction performance, IDBO still achieves further improvement on this basis, with RMSE reduced by about 8%, MAE reduced by about 7%, and R2 increased by nearly 0.6%. This proves the effectiveness of strategies such as chaotic initialization, adaptive step size, convex lens inverse imaging, and random difference mutation in the IDBO algorithm.

[0173] Table 4 Evaluation indicators of different optimization algorithms

[0174]

[0175] from Fig.12 It can be seen from the prediction curve that the prediction curve of the IDBO-Informer model is most consistent with the actual load data, especially the prediction at the peak and trough is more accurate. Fig.13 The absolute error graph also intuitively demonstrates the significant effect of the IDBO algorithm in reducing prediction errors, and its error fluctuation range is significantly smaller than that of other algorithms.

[0176] This paper compares the prediction performance of five modal decomposition methods (EMD, EEMD, CEEMD, CEEMDAN, and ICEEMDAN) combined with the Informer model. From the three key indicators of RMSE, MAE, and R2 score, the performance of each model is significantly different. Fig.14 As shown in Table 5, the EMD method, as the most basic method, has the highest RMSE and MAE values, which are 310.7 and 272.6 respectively, and the lowest R2 score, which is only 0.894, indicating that the prediction accuracy and fit are both low. With the improvement of the method, EEMD improves the decomposition effect by introducing white noise, and the RMSE and MAE are reduced to 279.2 and 236.3 respectively, and the R2 score is increased to 0.956, showing a significant performance improvement. CEEMD further optimizes EEMD. By integrating multiple EMD decomposition results, the RMSE and MAE are further reduced to 219.6 and 187.3, and the R2 score is also increased to 0.969. CEEMDAN adds a noise-assisted strategy based on CEEMD, making the decomposition more accurate, and the RMSE and MAE are reduced to 186.8 and 146.5 respectively, and the R2 score is increased to 0.971. However, as the latest improved method, ICEEMDAN significantly improves the decomposition accuracy by optimizing the introduction of adaptive noise and running EMD multiple times. Its RMSE and MAE drop sharply to 109.2 and 81.3, respectively, which is nearly 200 points lower than the EMD method. The R2 score is as high as 0.991, which is an increase of nearly 10%, indicating that its prediction accuracy and fit have reached extremely high levels.

[0177] Table 5 Evaluation indicators of different modal decomposition methods

[0178]

[0179]

[0180] Another specific embodiment of the present application provides a short-term load forecasting device based on the ICEEMDAN-PE and IDBO-Informer models, including a processor and a memory, wherein the memory stores a computer program, and when the computer program is called and executed by the processor, the method described above is implemented.

[0181] Another specific embodiment of the present application provides a computer-readable medium, wherein the computer-readable medium stores a computer program, and when the computer program is called and executed by a computer, the method described above is implemented.

[0182] The above is only a preferred implementation of the present application. It should be pointed out that a person skilled in the art can make several modifications and improvements without departing from the inventive concept of the present application, and these all fall within the scope of protection of the present application.

Claims

1. A short-term load forecasting method, characterized in that: The following steps are involved: S1: Use wavelet soft threshold denoising (WSTD) technology to remove noise from the original load data and improve data quality; S2: The denoised data is decomposed at multiple scales using the improved complete ensemble empirical mode decomposition (ICEEMDAN) method to obtain multiple intrinsic mode function (IMF) components and residual terms to capture the characteristics of different time scales. S3: The permutation entropy PE algorithm is used to evaluate the complexity of each IMF component; S4: Improve the Informer model and use the improved dung beetle optimization IDBO algorithm to optimize the parameters of the Informer model, including integrating chaos and reverse learning strategies for population initialization, introducing adaptive step size and convex lens inverse imaging strategy and random difference mutation strategy to enhance global search capabilities and avoid local optimality; S5: Input the decomposed data into the optimized Informer model for prediction, and integrate the prediction results of each IMF component according to the weighted PE value to obtain the final load prediction value.

2. The short-term load forecasting method according to claim 1, characterized in that: In step S1, WSTD is used for data preprocessing; the wavelet can be scaled and time-shifted by a scale factor a and a time-shift factor b to generate a series of sub-wavelets to extract target features at different resolutions; The specific steps of WSTD are as follows: The load sequence is denoised and modeled. It is assumed that the pure signal of the load sequence is f(t), and the sample signal of the load sequence containing noise is x(t): x(t)=f(t)+σn(t) (1) Where t is the time point; n(t) is the probability distribution that follows the normal distribution N(0,σ 2 )’s Gaussian white noise; In order to obtain the wavelet sequence equation using the dyadic wavelet, a and b are substituted into the wavelet sequence obtained by scaling and time shifting the wavelet: In the formula, ψ m,n (t) is the wavelet sequence; ψ(t) represents the wavelet function; In order to decompose the sample signal using DWT, a and b are incorporated into the CWT definition of wavelet ψ to obtain the corresponding DWT; Where W x Represents wavelet transform coefficients; Mallat algorithm is often used in DWT, and scaling functions and wavelet equations are often used as low-pass and high-pass filters; Mallat algorithm is used to perform multi-level decomposition of DWT; after decomposition, the corresponding scaling coefficients and wavelet coefficients are obtained; scaling coefficients contain low-frequency information, and wavelet coefficients contain high-frequency information: In the formula, c m,n is the scale factor; d m,n are wavelet coefficients; The threshold λ is calculated to distinguish the noise signal from the valid signal. According to the different threshold functions, WTD is divided into wavelet hard threshold denoising in equation (5) and WSTD in equation (6): Where W x is the wavelet transform coefficient; λ is the threshold; |W x | is the evaluated wavelet transform coefficient; the noise wavelet transform coefficient is set to zero and the load time series is reconstructed.

3. The short-term load forecasting method according to claim 1, characterized in that: In step S2: CEEMDAN introduces adaptive noise and runs EMD multiple times to obtain multiple groups of modal functions, which are then averaged to obtain a more accurate decomposition result. The formula is as follows: In the formula, ω (i) is Gaussian white noise; r k is the kth group of residuals; I is the generated ω (i) ; M(·) is the local mean of the generated sequence.

4. The short-term load forecasting method according to claim 1, characterized in that: In step S3: The PE algorithm is used to measure the complexity of each sub-module component; The steps of PE algorithm are as follows: (1) The reconstructed sequence of the phase space of the time series can be obtained by the following formula: Y s =[x(s),x(s+τ),…x(s+(m-1)τ)](9) Where s = 1, 2, ..., N-(m-1)τ, τ is the delay time, and m is the embedding dimension; (2) For each reconstructed vector Y s Rearrange the components of from small to large to get the symbol sequence: Y(g)=(s1,s2,…,s m ),g=1,2,…,landl≤m!(10) (3) The probability of each symbol sequence Y(g) occurring is calculated as P g : Where N(Y(g)) is the number of Y(g), P g satisfy (4) The PE of this time series is defined as: From formula (12), we can see that PE is an indicator to measure the complexity of a sequence. The larger the PE value, the more complex the sequence. The complexity of each subsequence is evaluated by calculating the PE value of each IMF.

5. The short-term load forecasting method according to claim 1, characterized in that: In step S4: The improvements of Informer include: using the probabilistic sparse self-attention mechanism ProbSparse to reduce the complexity of matrix calculation; introducing the self-attention distillation mechanism to extract the main features of the time series; using the decoder to directly generate output prediction values; the formula of ProbSparse self-attention is: Where: is a sparse matrix of the same size as Q; The self-attention distillation operation is introduced, and the convolution and pooling operations are used to reduce the dimension and network parameters; the distillation process from the jth layer to the j+1th layer is as follows: In the formula, [·] AB represents the attention block, which contains multi-head ProbSparse self-attention and basic operations; Conv1d(·) uses the ELU activation function to perform one-dimensional convolution filtering on the time dimension; represents downsampling; The input of the decoder is represented as: In the formula, Start token; Placeholder for the target sequence that sets the scalar to 0; L token is the length of the input; L y is the length of the output sequence; d model is the dimension of the model input; in the ProbSparse self-attention calculation, masked multi-head attention is applied by setting the masked dot product to -∞.

6. The short-term load forecasting method according to claim 1, characterized in that: In step S4: The IDBO algorithm is an improvement on the DBO algorithm. The IDBO algorithm combines the chaos initialization method with the reverse learning initialization strategy to improve the convergence speed of DBO. The specific steps are as follows: Fuch mapping is used to generate the initial population of DBO. Where P(t) is the chaotic variable of Fuch mapping. When the Fuch chaotic mapping is used to generate the initial solution, the reverse learning strategy is introduced. The mathematical expression of the reverse learning strategy is shown in formula (17): ON i =K(P min +P max )-P i (17) In the formula, OP i is the inverse solution P corresponding to each initial solution i ;P min and P max Respectively represent the minimum and maximum values ​​in the initial solution; K is a random value in the range of (0,1); A dynamic selection strategy is adopted to update the target position alternately with the adaptive step size strategy and the convex lens imaging inversion strategy under a certain probability; the strategy is mainly determined by the linearly decreasing adaptive step control factor α0, as follows: At the same time, the convex lens imaging learning strategy is introduced to disturb the foraging dung beetle population. The equation is: Among them, P j Represents the component of the current individual in dimension j; a j and b j They represent the minimum and maximum values ​​of the j dimension in the search space respectively; Expressing the inverse solution of the convex lens P j ; Then, the choice of which strategy to use for the target position update is determined by the selection probability P s Determine, as shown in formula (20); if the random value P s <0.5, the random stepping strategy is used to update the position of the foraging dung beetle; otherwise, the convex lens reverse learning strategy is used to update the position, as shown in formula (21): A random difference mutation strategy was introduced to enhance the population diversity of dung beetles; the equation is as follows: P(t+1)=r1×(P * -P(t))+r2×(P′(t)-P(t)) (22) Where P(t+1) is a new individual obtained by random differential mutation, r1 and r2 are random values ​​in the range of [0, 1]; * is the current optimal individual position; P′(t) is the position of a randomly selected individual in the population.

7. A short-term load forecasting device, characterized in that: The method comprises a processor and a memory, wherein the memory stores a computer program, and when the computer program is called and executed by the processor, the method according to any one of claims 1 to 6 is implemented.

8. A computer-readable medium, characterized in that The computer-readable medium stores a computer program, and when the computer program is called and executed by a computer, the method according to any one of claims 1 to 6 is implemented.

Images