Short-term power load prediction method and system based on optimization integration

Abstract

The invention relates to a short-term power load prediction method and system based on optimization integration. The method comprises the following steps: acquiring original power load data; carrying out abnormal value detection, missing value detection and data filling on the power load data set, and constructing an input sequence; performing season trend decomposition on the input sequence, and extracting a season component, a trend component and a fluctuation component; an Informer model is trained, and prediction of each component is completed; constructing a weight optimization model by taking the prediction result of each component as input and taking the sum of the mean absolute error and the standard deviation of the prediction error as an optimization target; and optimizing the weight of each component through an IDBO algorithm, and carrying out weighted summation on each component according to an optimization result to obtain a final short-term power load prediction result. According to the method, the season, trend and fluctuation components of the power load are extracted, the corresponding Informer model is constructed, and weighted summation is carried out on the prediction results of the components by adopting an integrated optimization idea, so that the power load prediction precision and stability are effectively improved.

Description

Technical Field

[0001] The present invention relates to the field of power load forecasting, and in particular to a short-term power load forecasting method and system based on optimization integration. Background Art

[0002] Load forecasting is a key link in the power system and has a significant impact on the safe and economic operation of the power system. With the development of the economy and the progress of society, the demand for electricity is increasing, and the load characteristics are becoming more and more complex and diverse. Accurate load forecasting can effectively improve the operating efficiency of the power system, reduce operating costs, and ensure the stability and reliability of power supply.

[0003] A power load forecasting method based on secondary decomposition is proposed in the prior art, which can effectively extract data information, but the modal decomposition number of VMD needs to be set in advance, and an unreasonable modal decomposition number is likely to lead to modal redundancy or loss of important modal information; a short-term power load forecasting method and system based on artificial intelligence, whose prediction model adopts LSTM, but LSTM has the phenomenon of information loss when facing long time series. Summary of the invention

[0004] The purpose of the present invention is to solve the above problems and propose a short-term power load forecasting method and system based on optimization integration. The seasonal trend decomposition method (Seasonal and Trend decomposition using Loess, STL) is used to extract the seasonal, trend and fluctuation characteristics of the power load, which are used to train the Informer model respectively. The integrated optimization idea is used to perform weighted summation of the prediction results of each component, so as to effectively improve the accuracy and stability of power load forecasting.

[0005] In order to achieve the above object, the technical solution provided by the present invention is: A short-term power load forecasting method and system based on optimization integration, comprising: S1. Obtaining original power load sequence data; S2, preprocessing the power load sequence data, wherein the preprocessing includes outlier detection, missing value detection and data filling, and constructing a power load data set using the preprocessed power load sequence data; S3, decomposing the power load data set using a seasonal trend decomposition method, and extracting the seasonal component, trend component, and fluctuation component of the power load data set; S4. Divide the seasonal component, trend component and fluctuation component of the power load data set into a training set and a test set respectively in proportion, and construct corresponding informer models for each component as seasonal component prediction model, trend component prediction model and fluctuation component prediction model, and after training and testing respectively, use the corresponding informer model to obtain the predicted values ​​of seasonal component, trend component and fluctuation component; S5. Taking the forecast values ​​of seasonal component, trend component and fluctuation component as input, taking the sum of mean absolute error and forecast error standard deviation as optimization target, constructing a weight optimization model for forecast component integration, which is used to solve the optimal weights of forecast values ​​of seasonal component, trend component and fluctuation component; S6. Introducing multiple strategies into the dung beetle optimization algorithm to obtain an improved dung beetle optimization algorithm, using the improved dung beetle optimization algorithm to solve the optimization model of the prediction component integration, and optimizing the weight coefficients of the prediction values ​​of the seasonal component, the trend component and the fluctuation component; S7. According to the optimal weight coefficients of the predicted values ​​of the seasonal component, trend component and fluctuation component obtained in step S6, the real-time predicted values ​​of the seasonal component, trend component and fluctuation component output by the seasonal component prediction model, trend prediction model and fluctuation prediction model are weighted summed to obtain the short-term power load forecast result.

[0006] Furthermore, in S1, the power load data of a certain region is collected, the sampling frequency remains unchanged, and the time period is one year; Furthermore, in S2, the box plot theory is used to test the data anomalies and missing data caused by equipment factors and human factors during the data collection process; the linear interpolation method is used to fill the abnormal values ​​and missing values ​​of the power load that appear in the detection process. The calculation formula is as follows:

[0007] In the formula, x t for t Abnormal or missing data at a certain moment, t Indicates missing time, x t+1 is the moment after the missing moment, x t-1 The data is the previous moment of the missing moment; Furthermore, in S3, the input sequence is decomposed using STL to extract the trend component, seasonal component and fluctuation component that characterize the change characteristics of the power load, including the following steps: 1. Internal circulation: 1) Using the estimated trend component For the original sequence Y tDetrending is performed, and the calculation formula is: ; In the formula, To estimate the trend component; Y t is the original sequence; is the detrended component; 2) Based on the local weighted regression method, the subsequences of each period after removing the trend are smoothed to obtain the initial seasonal component ; 3) Use low-pass filtering and local weighted regression methods to process the initial seasonal component , and obtain the residual trend component ; 4) Calculate the difference between the initial seasonal component and the residual trend quantity to obtain the seasonal component:

[0008] In the formula, For seasonal components; is the initial seasonal component; is the residual trend component; 5) Remove the original sequence Y t Seasonal components in:

[0009] In the formula, To go to the seasonal portion; For seasonal components; Y t is the original sequence; 6) Apply non-parametric local weighted regression smoothing to the seasonal component in step 5) to obtain the trend component .

[0010] (II) External circulation: According to the trend component and seasonal component, the volatility component is calculated, and the original sequence is decomposed into seasonal component, trend component and volatility component. The calculation formula is:

[0011] In the formula, Y t is a time series; T t Represents the trend component; S t Represents seasonal component; R t Represents the fluctuation component.

[0012] Furthermore, in S4, the seasonal component, trend component and fluctuation component are normalized respectively, as follows:

[0013] In the formula, f n is the normalized data of each component; f is the current data; f max and f min are the maximum and minimum values ​​of the sequence respectively.

[0014] Furthermore, in S4, the normalized sequence is divided into a training set and a test set in a ratio of 8:2, and the Informer model is trained with the training set and the prediction results of the seasonal component, trend component and fluctuation component are obtained. The Informer model includes: In the encoder, multi-head probabilistic sparse self-attention is used to extract time series feature information, and the expression is as follows:

[0015] In the formula, Q Represents the Query matrix; K represents the Key matrix; V Represents the Value matrix; represent Q The probability coefficient matrix of ; d k Represents the input dimension; Softmax represents the activation function; A It is the self-attention mechanism; T is transposed.

[0016] The distillation mechanism is further used to give high weights to important features and to focus more on them in the next layer of sparse self-attention, which effectively avoids the problem of feature redundancy. j Layer to j The distillation process of the +1 layer is calculated as follows:

[0017] In the formula, j Represents the number of distillation layers; X represents the input sequence of embeddings; represents the attention block; Convld Represents one-dimensional convolution filtering; MaxPool It is the maximum pooling operation; ELU is the activation function.

[0018] The decoder consists of multi-head self-attention and masked multi-head probabilistic sparse self-attention. The decoder uses multi-head self-attention to interact with the encoded features. The masked multi-head probabilistic sparse self-attention masks future data information. Its input vector expression is as follows:

[0019] In the formula, The decoder t An input sequence; For the t The start mark of a sequence; For the t Target sequence placeholder for sequences, Concat For the connector.

[0020] Preferably, the fully connected layer is used to change the dimension of the output data to achieve a one-time output of the final prediction result of the time series.

[0021] Furthermore, in S5, the load forecasting method needs to take into account both forecasting accuracy and stability, and the mean absolute error M AE Can characterize prediction accuracy and prediction error standard deviation S D The prediction stability can be characterized, and the sum of the two is taken as the optimization target, and the weight coefficient of each component is taken as the decision variable to construct a weight optimization model, as follows:

[0022] In the formula, c 1. c 2 and c 3 are the weight coefficients corresponding to the trend series, seasonal series and fluctuation series respectively; is the actual load; To predict load; , and They are the forecast results of trend series, seasonal series and fluctuation series respectively; is the mean value of load forecast error; N is the sample size; M AE is the mean absolute error; S D is the standard deviation; F is the objective function.

[0023] Furthermore, in S6, multiple strategies are used to improve DBO, which can effectively improve the global search capability and convergence speed, including: 1) Use the infinitely foldable chaotic map Fuch to map the variables. The calculation formula is as follows:

[0024] Where: P t is the chaotic variable of the Fuch map, P t+1 is the new individual after mapping; 2) The reverse learning strategy is conducive to expanding the search space of dung beetles and improving the quality of the initial solution of the population. The calculation formula is as follows:

[0025] Where: OP i For the initial solution P i The inverse solution of P max and P min are the maximum and minimum values ​​in the initial solution respectively; K is a random number in the interval [0,1]; 3) Update the target position alternately based on the adaptive step size strategy and the convex lens inverse imaging strategy, as follows: Determine the target position update strategy based on the selection probability, and the calculation formula is:

[0026] In the formula, P s represents the probability of selection, t and T are the current number of iterations and the maximum number of iterations respectively; like P s <0.5, the adaptive step size strategy is used to update the position of the foraging dung beetle; otherwise, the convex lens inverse imaging strategy is used to update the position, and the calculation formula is as follows:

[0027] Where: Indicated in t In the iteration i The location of the individual; Indicated in t +1 iteration i The location of the individual; is the adaptive step control factor; c 1 is a random number that satisfies the normal distribution; c 2 represents a random vector in the interval [0,1]; Represents the current global optimal value; and They are the upper and lower limits of the foraging area of ​​juvenile dung beetles; krepresents the deflection coefficient; 4) Introduce a random difference mutation strategy to enhance the population diversity of thieving dung beetles. The calculation formula is:

[0028] Where: P t+1 New individuals obtained by random differential mutation; For the current individual; r 1 and r 2 are all random numbers in the interval [0,1]; is the best position information of the current individual; is the position of a randomly selected individual in the population.

[0029] Furthermore, the mean absolute error M AE and the standard deviation of the prediction error S D The sum is the optimization objective function, and the weight coefficient c 1. c 2 and c 3 is the decision variable, solve the weight optimization model, and get the weight coefficient corresponding to each component c 1. c 2 and c 3.

[0030] Further, in S7, according to the weight coefficient c 1. c 2 and c 3. Perform weighted summation of trend component, seasonal component and fluctuation component to complete short-term power load forecast.

[0031] Compared with the prior art, the present invention has the following technical effects: 1) The present invention extracts power load sequence information through seasonal trend decomposition method, further adopts Informer to predict each component separately, and fully utilizes the differences of each component by constructing a weight optimization model.

[0032] 2) The box plot method is used to detect outliers and analyze data distribution characteristics of power load sequence data, and the linear interpolation method is used to fill in the outliers and missing values ​​in the detection process, thereby improving the robustness of the load forecasting method of the present invention. 3) The present invention comprehensively considers the prediction accuracy and stability of the model, constructs a weight optimization model to achieve the optimal integration of seasonal components, trend components and fluctuation components, and effectively improves the load prediction accuracy and stability.

[0033] 4) The present invention solves the weight optimization model through an improved dung beetle optimization algorithm, which has a higher convergence speed and global convergence ability. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] The present invention will be further described below in conjunction with the accompanying drawings and embodiments: Figure 1 This is a flow chart of power load forecasting according to an embodiment of the present invention; Figure 2 A schematic diagram of a power load forecasting structure flow chart of an embodiment of the present invention; Figure 3 Schematic diagram of comparison between predicted values ​​and true values ​​of each model in an embodiment of the present invention; Figure 4 4 is a system module flow chart of an embodiment of the present invention. DETAILED DESCRIPTION

[0035] To make the purpose, technical solution and advantages of the embodiments of the present invention more clear, the technical solution of the present invention will be described in detail below in conjunction with the accompanying drawings in the embodiments of the present invention.

[0036] Example 1 like Figure 1 As shown, the short-term power load forecasting method based on optimization integration described in the present invention comprises the following steps: S1. Acquisition of original power load data set; S2, performing outlier detection, missing value detection and data filling on the power load data set, and constructing an input sequence using the pre-processed power load data set; S3, performing seasonal trend decomposition on the input sequence to extract seasonal component, trend component and fluctuation component; S4, dividing each component into a training set and a test set according to a fixed ratio, and training an Informer model with each component as input; S5, using the prediction results of each component as input, taking the sum of the mean absolute error and the standard deviation of the prediction error as the optimization target, and constructing a weight optimization model; S6, optimizing the weight coefficients of each component by improving the DBO algorithm through multiple strategies; S7. Perform weighted summation on each component according to the optimization result to obtain a final short-term power load forecast result.

[0037] like Figure 2 As shown, in this embodiment, the STL decomposition includes two parts: an inner loop and an outer loop, which are as follows: The seasonal smoothing and subsequent trend smoothing in the inner loop are used to update the seasonal component and the trend component respectively, including the following steps: 1) Using the estimated trend component For the original sequence Y tDetrending is performed and the calculation is as follows:

[0038] In the formula, To estimate the trend component; Y t is the original sequence; is the detrended component; 2) Based on local weighted regression, smooth the subsequences of each period after removing the trend to obtain the initial seasonal component ; 3) Processing the initial seasonal component using low-pass filtering and local weighted regression , and obtain the residual trend component ; 4) Calculate the difference between the initial seasonal component and the residual trend quantity to obtain the seasonal component:

[0039] In the formula, For seasonal components; is the initial seasonal component; is the residual trend component; 5) Remove the original sequence Y t Seasonal components in:

[0040] In the formula, To go to the seasonal portion; For seasonal components; Y t is the original sequence.

[0041] 6) Apply non-parametric local weighted regression smoothing to the seasonal component in step 5 to obtain the trend component .

[0042] In the outer loop, the volatility component is calculated based on the trend component and the seasonal component. The original sequence is decomposed into seasonal component, trend component and volatility component as follows:

[0043] In the formula, Y t is a time series; T t Represents the trend component; S t Represents seasonal component; R t Represents the fluctuation component.

[0044] In this embodiment, the Informer model includes two parts: an encoder and a decoder, as follows: In the encoder, multi-head probabilistic sparse self-attention is used to extract time series feature information. Compared with the traditional self-attention mechanism, its advantage is to reduce computational complexity and storage space. The specific expression is as follows:

[0045] In the formula, Q Represents the Query matrix; K represents the Key matrix; V Represents the Value matrix; represent Q The probability coefficient matrix of ; d k Represents the input dimension; Softmax represents the activation function; A It is the self-attention mechanism; T is transposed.

[0046] The distillation mechanism is further used to give high weights to important features and to focus more on them in the next layer of sparse self-attention, which effectively avoids the problem of feature redundancy. j Layer to j The distillation process of the +1 layer is calculated as follows:

[0047] In the formula, j Represents the number of distillation layers; X represents the input sequence of embeddings; represents the attention block; Convld Represents one-dimensional convolution filtering; MaxPool It is the maximum pooling operation; ELU is the activation function.

[0048] The decoder consists of multi-head self-attention and masked multi-head probabilistic sparse self-attention. The former is used to interact with the encoded features, and the latter avoids the autoregression problem of the model by masking future data information. Its input vector expression is as follows:

[0049] In the formula, The decoder t An input sequence; For the t The start mark of a sequence; For the t Target sequence placeholder for sequences; Concat For the connector.

[0050] The fully connected layer is used to change the dimension of the output data so that the final prediction results of the time series can be output at one time.

[0051] In this embodiment, the weight optimization model is as follows:

[0052] In the formula, c 1. c 2 and c 3 are the weight coefficients corresponding to the trend series, seasonal series and fluctuation series respectively; is the actual load; To predict load; , and They are the forecast results of trend series, seasonal series and fluctuation series respectively; is the mean value of load forecast error; N is the sample size; M AE is the mean absolute error; S D is the standard deviation; F is the objective function.

[0053] In this embodiment, the multi-strategy improved DBO algorithm is specifically as follows: Population initialization based on fusion chaos and reverse learning strategy: The dung beetle optimization algorithm initializes the population by random generation, which is prone to low coverage of the solution space, which is not conducive to algorithm optimization. Fuch is an infinitely foldable chaotic map with good ergodicity, dynamics and convergence, as follows:

[0054] Where: P t is the chaotic variable of the Fuch map; P t+1 is the new individual after mapping.

[0055] The reverse learning strategy is beneficial to expand the search space of dung beetles and improve the quality of the initial solution of the population, as follows:

[0056] Where: P i is the initial solution; OP i for P i The inverse solution of P max and P min are the maximum and minimum values ​​in the initial solution respectively;K is a random number in the interval [0,1].

[0057] Integration of adaptive step size and convex lens inverse imaging strategy: In view of the lack of adaptive ability and weak global search ability of dung beetles in the foraging process, an adaptive step size strategy and a convex lens inverse imaging strategy are introduced to alternately update the target position. In the initial iteration, a large step size can expand the foraging search space, improve the global search ability of the algorithm, and accelerate the convergence speed; in the later iteration, a small step size is conducive to the local search of the algorithm. Therefore, according to the selection probability P s To determine the target position update strategy, as follows:

[0058] In the formula, t and T are the current number of iterations and the maximum number of iterations respectively.

[0059] like P s <0.5, the adaptive step size strategy is used to update the position of the foraging dung beetle; otherwise, the convex lens inverse imaging strategy is used to update the position, which is calculated as follows:

[0060] Where: Indicated in t In the iteration i The location of the individual; P t+1 For the new individuals obtained; is the adaptive step control factor; c 1 is a random number that satisfies the normal distribution; c 2 represents a random vector in the interval [0,1]; Represents the current global optimal value; and They are the upper and lower limits of the foraging area of ​​juvenile dung beetles; k is the deflection coefficient; P s is the probability of selection.

[0061] Random difference mutation strategy: The position update strategy of the thieving dung beetle is based on the current optimal state of the individual, which can easily cause the algorithm to fall into a local optimum. Therefore, the random difference mutation strategy is introduced to enhance the population diversity of the thieving dung beetle. The calculation formula is:

[0062] Where: P t+1 For the new individuals obtained; r 1 andr 2 are all random numbers in the interval [0,1]; is the best position information of the current individual; is the position of a randomly selected individual in the population; Indicated in t In the iteration i The location of the individual.

[0063] In order to verify the effectiveness of the present invention, an example analysis is conducted using the power load of a region in Australia in 2010. The specific time is from 00:00 on January 1, 2010 to 23:30 on December 31, 2010. The sampling frequency is 30 minutes, the total number of samples is 17520, and the determination coefficient, mean absolute error and standard deviation are used as evaluation indicators, as follows:

[0064]

[0065]

[0066] In the formula, N is the sample size; is the actual value; is the predicted value; Indicates the load average value; is the mean load forecast error. R 2 is the determination coefficient.

[0067] The model prediction comparison results are shown in Table 1:

[0068] If the original load data is directly input into the four single models, each model will find it difficult to capture the trend characteristics, cycle characteristics and fluctuation characteristics implied by the power load, which will have a greater impact on the load forecasting accuracy and stability. R 2 They increased by 6.76%, 4.85%, 2.26% and 2.56% respectively; M AE They decreased by 74.04%, 69.51%, 59.74% and 61.37% respectively; S D They decreased by 63.84%, 59.44%, 49.68% and 50.19% respectively.

[0069] like Figure 3 As shown, the prediction results of the present invention are most closely matched to the true values, while there are large deviations between the prediction results of the other four single models and the true values.

[0070] Example 2 like Figure 4 As shown, this embodiment discloses a short-term power load forecasting system based on optimization integration, including an acquisition module, a preprocessing module, a decomposition module, a prediction module, an integrated weight optimization module and a load output module, as follows: Collection module: used to collect power load data; Preprocessing module: used to perform abnormal value detection, missing value detection, and data filling on the collected power load data, and construct a power load input sequence using the preprocessed power load data; Decomposition module: used to decompose the power load input sequence and extract the trend information, seasonal information and fluctuation information of the power load data; Prediction module: used to normalize the decomposition results, divide the training set and the test set according to a fixed ratio, and input them into the Informer model to obtain the prediction results of each component; Integrated weight optimization module: used to input the forecast results of trend component, seasonal component and fluctuation component into the weight optimization model, and obtain the optimal weight coefficient corresponding to each component through the multi-strategy improved DBO algorithm; Load output module: used to perform weighted summation of each component to obtain the final short-term power load output result.

[0071] It is particularly noted that the above examples are only some of the better examples of the present invention. The present invention is described in detail with reference to the above embodiments. Those skilled in the art should understand that the specific implementation modes of the present invention may be modified or replaced by equivalents, and any modification or equivalent replacement that does not depart from the spirit and scope of the present invention shall be included in the scope of protection of the claims of the present invention.

Claims

1. A short-term power load forecasting method and system based on optimization integration, characterized in that: The following steps are involved: S1. Obtaining original power load sequence data; S2, preprocessing the power load sequence data, wherein the preprocessing includes outlier detection, missing value detection and data filling, and constructing a power load data set using the preprocessed power load sequence data; S3, decomposing the power load data set using a seasonal trend decomposition method, and extracting the seasonal component, trend component, and fluctuation component of the power load data set; S4. Divide the seasonal component, trend component and fluctuation component of the power load data set into a training set and a test set respectively in proportion, and construct corresponding informer models for each component as seasonal component prediction model, trend component prediction model and fluctuation component prediction model, and after training and testing respectively, use the corresponding informer model to obtain the predicted values ​​of seasonal component, trend component and fluctuation component; S5. Taking the forecast values ​​of seasonal component, trend component and fluctuation component as input, taking the sum of mean absolute error and forecast error standard deviation as optimization target, constructing a weight optimization model for forecast component integration, which is used to solve the optimal weights of forecast values ​​of seasonal component, trend component and fluctuation component; S6, introducing multiple strategies into the dung beetle optimization algorithm to obtain an improved dung beetle optimization algorithm, using the improved dung beetle optimization algorithm to solve the weight optimization model of the prediction component integration, and optimizing the weight coefficients of the prediction values ​​of the seasonal component, the trend component and the fluctuation component; S7. According to the optimal weight coefficients of the predicted values ​​of the seasonal component, trend component and fluctuation component obtained in step S6, the real-time predicted values ​​of the seasonal component, trend component and fluctuation component output by the seasonal component prediction model, the trend component prediction model and the fluctuation component prediction model are weighted summed to obtain the short-term power load forecast result.

2. The short-term power load forecasting method according to claim 1, characterized in that: In step S2, the box plot method is used to perform outlier detection and data distribution feature analysis on the power load sequence data, and the linear interpolation method is used to fill the outliers and missing values ​​in the detection process. The calculation formula is: In the formula, x t for t Abnormal or missing data at a certain moment, t Indicates missing time, x t+1 is the moment after the missing moment, x t-1 The data is the previous moment of the missing moment.

3. The short-term power load forecasting method according to claim 2, characterized in that: In the step S3, a seasonal trend decomposition method is used to extract a trend component, a seasonal component and a fluctuation component that characterize the change characteristics of the power load; The seasonal trend decomposition method includes two parts: internal cycle and external cycle; Seasonal smoothing and subsequent trend smoothing in the inner loop are used to update the seasonal component and trend component respectively, including: 1) Using the estimated trend component For the original sequence Y t Detrending is performed, and the calculation formula is: ； In the formula, To estimate the trend component; Y t is the original sequence; is the detrended component; 2) Based on the local weighted regression method, the subsequences of each period after removing the trend are smoothed to obtain the initial seasonal component ; 3) Use low-pass filtering and local weighted regression methods to process the initial seasonal component , and obtain the residual trend component ; 4) Calculate the difference between the initial seasonal component and the residual trend quantity to obtain the seasonal component: In the formula, For seasonal components; is the initial seasonal component; is the residual trend component; 5) Remove the original sequence Y t Seasonal components in: In the formula, To go to the seasonal portion; For seasonal components; Y t is the original sequence; 6) Apply non-parametric local weighted regression smoothing to the seasonal component in step 5) to obtain the trend component ; In the external loop, the volatility component is calculated based on the trend component and the seasonal component, and the original sequence is decomposed into the seasonal component, the trend component and the volatility component. The calculation formula is: In the formula, Y t is a time series; T t Represents the trend component; S t Represents seasonal component; R t Represents the fluctuation component.

4. The short-term power load forecasting method according to claim 3 is characterized in that: In step S4, the trend component, seasonal component and fluctuation component obtained in step S3 are normalized, the normalized data sequence is divided into a training set and a test set according to a fixed ratio, and the Informer model is input and trained to obtain the prediction results of the trend component, seasonal component and fluctuation component.

5. The short-term power load forecasting method according to claim 4, characterized in that: In step S5, the sum of the mean absolute error and the error standard deviation is used as the optimization target, and the weight coefficients of the trend component, the seasonal component and the fluctuation component are used as decision variables to construct a weight optimization model, which is specifically as follows: In the formula, γ 1. γ 2 and γ 3 are the weight coefficients corresponding to the trend series, seasonal series and fluctuation series respectively; is the actual load; To predict load; , and They are the forecast results of trend series, seasonal series and fluctuation series respectively; is the mean value of load forecast error; N is the sample size; M AE is the mean absolute error; S D is the standard deviation; F is the objective function.

6. The short-term power load forecasting method according to claim 5, characterized in that: In step S6, the improved dung beetle optimization algorithm specifically includes: 1) Use the infinitely foldable chaotic map Fuch to map the variables. The calculation formula is as follows: Where: P t is the chaotic variable of the Fuch map, P t+1 is the new individual after mapping; cos is the cosine function; 2) Use the reverse learning strategy to expand the search space of dung beetles and improve the quality of the initial solution of the population. The calculation formula is as follows: Where: OP i For the initial solution P i The inverse solution of P max and P min are the maximum and minimum values ​​in the initial solution respectively; K is a random number in the interval [0,1]; 3) Update the target position alternately based on the adaptive step size strategy and the convex lens inverse imaging strategy, as follows: Determine the target position update strategy based on the selection probability, and the calculation formula is: In the formula, P s represents the probability of selection, t and T are the current number of iterations and the maximum number of iterations respectively; like P s <0.5, the adaptive step size strategy is used to update the position of the foraging dung beetle; otherwise, the convex lens inverse imaging strategy is used to update the position, and the calculation formula is: Where: Indicated in t In the iteration i The location of the individual; Indicated in t +1 iteration i The location of the individual; is the adaptive step control factor; c 1 is a random number that satisfies the normal distribution; c 2 represents a random vector in the interval [0,1]; Represents the current global optimal value; , are the upper and lower limits of the foraging area of ​​juvenile dung beetles; k represents the deflection coefficient; 4) Introduce a random difference mutation strategy to enhance the population diversity of thieving dung beetles. The calculation formula is: Where: P t+1 New individuals obtained by random differential mutation; For the current individual; r 1 and r 2 are all random numbers in the interval [0,1]; is the best position information of the current individual; is the position of a randomly selected individual in the population.

7. A system for the short-term power load forecasting method according to any one of claims 1 to 6, characterized in that: include: Collection module: used to collect power load data; Preprocessing module: used to perform outlier detection, missing value detection and data filling on the collected power load data, and construct a power load input sequence using the preprocessed power load data; Decomposition module: used to decompose the power load input sequence and extract the trend information, seasonal information and fluctuation information of the power load data; Prediction module: used to normalize the decomposition results, divide the training set and the test set according to a fixed ratio, and input them into the Informer model to obtain the prediction results of each component; An integrated weight optimization module is used to input the prediction results of each component into the weight optimization model, and obtain the optimal weight coefficient corresponding to each component through the multi-strategy improved DBO algorithm; The load output module is used to perform weighted summation of each component to obtain the final short-term power load output result.

Images