Short-term power load prediction method based on hybrid model

Abstract

The invention belongs to the technical field of power system load prediction, and particularly relates to a short-term power load prediction method based on a hybrid model, and the method comprises the steps: S1, carrying out the preprocessing of original data; s2, searching an optimal combination [k, alpha] by adopting a firefly and sparrow algorithm; s3, decomposing the original data by adopting variational mode decomposition; s4, determining a demarcation point by adopting an adaptive algorithm, and dividing each signal component; s5, selecting features by adopting the maximum information coefficient, and establishing an input matrix; s6, substituting each signal component into the model for prediction; and S7, reconstructing the prediction result, and outputting a final prediction value. According to the method, a combined prediction model based on feature screening and frequency division prediction is provided, signal components are extracted from original data, the signal components are separated into a high-frequency part and a low-frequency part, after feature screening is conducted on the signal components, the signal components are substituted into the model for prediction, and finally all signal component prediction results are reconstructed. And obtaining a final prediction value.

Description

Technical Field

[0001] The invention belongs to the technical field of power system load forecasting, and specifically is a short-term power load forecasting method based on a hybrid model. Background Art

[0002] At present, electricity demand is affected by many factors such as weather, electricity prices and other unexpected events, and has a certain random volatility, which causes a certain fluctuation in power load. According to the length of the forecast time, power load forecasting is divided into long-term load forecasting, medium-term load forecasting and short-term load forecasting. Accurate short-term load forecasting can not only help the system run normally and avoid resource waste, but also improve economic benefits.

[0003] There are two types of methods for power load forecasting: one is machine learning methods, such as exponential smoothing, autoregressive integral moving average model, multivariate linear regression, Kalman filtering algorithm, grey prediction theory and support vector machine. These methods have low overall prediction accuracy, are difficult to use for large-scale data, and have poor universality; the other is deep learning prediction methods, which mainly include deep learning methods such as error back propagation algorithm, recurrent neural network, long short-term memory neural network, gated recurrent network and Informer model for load forecasting. This type of method can deeply understand the internal characteristics of the data and has strong nonlinear processing capabilities. However, in the context of large power grids, it has weak generalization ability and is easy to fall into local extreme values. The prediction accuracy is not high enough, and most of them only predict a single load without considering factors such as weather and electricity prices. The prediction time is long for multiple inputs.

[0004] Therefore, how to provide a short-term power load forecasting method is an urgent problem to be solved by those skilled in the art. Summary of the invention

[0005] In view of the above problems, the present invention provides a short-term power load forecasting method, which is based on a combined forecasting model of feature screening and frequency division forecasting. In order to extract signal components containing rich feature information from the original load sequence with more noise and achieve the best signal processing effect, the firefly sparrow algorithm is first used to optimize the influencing parameter combination in the variational mode decomposition to obtain the signal component with the best effect. At the same time, an adaptive algorithm is used to determine the high- and low-frequency zero-crossing rate separation point, and each signal component is divided into high-frequency and low-frequency parts. Secondly, the maximum information coefficient is used to perform feature screening on each signal component. Finally, the high-frequency signal component and its features are brought into the Informer model for prediction, and the low-frequency signal component and its features are brought into the LSTM model for prediction. Finally, the prediction results of all signal components are reconstructed to obtain the final prediction value.

[0006] The technical solution adopted by the present invention to solve the technical problem is: a short-term power load forecasting method based on a hybrid model described in the present invention comprises the following steps:

[0007] S1: preprocess the raw data;

[0008] S2: Use the firefly sparrow algorithm to optimize the parameter combination [k, α] in the variational mode decomposition and determine the best parameter combination [k, α];

[0009] S3: Decomposing the preprocessed raw power load data using variational mode decomposition;

[0010] S4: using an adaptive algorithm to determine the dividing point between high-frequency and low-frequency signal components, and dividing the decomposed power load data into a high-frequency signal component and a low-frequency signal component according to the frequency dividing point;

[0011] S5: using the maximum information coefficient to select 3 to 5 features with the highest correlation for each signal component from among the relevant features, and establishing an input matrix;

[0012] S6: Input the low-frequency signal input matrix into the LSTM model for prediction, and output the prediction result of the LSTM model on the low-frequency signal input matrix; input the high-frequency signal input matrix into the Informer model for prediction, and output the prediction result of the Informer model on the high-frequency signal input matrix;

[0013] S7: The prediction results of the LSTM model on the low-frequency signal input matrix and the prediction results of the Informer model on the high-frequency signal input matrix are superimposed and reorganized to output the final power load prediction result;

[0014] The original data includes but is not limited to original power load data, real-time electricity prices and weather factors, and the weather factors include but are not limited to temperature, humidity, rainfall, wind speed, maximum temperature, minimum temperature, average temperature, and season;

[0015] The original data are classified according to different administrative divisions, including provincial, municipal, county and township levels. After the original data are classified, they are respectively input into the model for prediction and multiple groups of prediction results are given.

[0016] Preferably, the raw data preprocessing in step S1 includes the following:

[0017] S1.1 uses linear interpolation to process outliers and missing values ​​and correct load data;

[0018] S1.2 In order to avoid the different dimensions of each eigenvalue, which affects the prediction accuracy and speeds up the gradient descent, the data is normalized uniformly:

[0019]

[0020] Among them, G' is the normalized data value; G is the original data value; G max and G min are the maximum and minimum values ​​in the original data set;

[0021] S1.3 Observe the daily average load curve of the original data and determine whether it has seasonality, periodicity, etc.

[0022] Preferably, the firefly sparrow algorithm in step S2 includes the following steps:

[0023] S2.1 Initialization parameters: set the population size, number of iterations and dimension, initialize the ratio of predators and joiners, calculate the fitness value, and sort;

[0024] S2.2 Sparrow updates predator location;

[0025] S2.3 Sparrow updates the joiner's position;

[0026] S2.4 Sparrow updates the position of the alerter;

[0027] S2.5 calculates the fitness value and updates the sparrow position;

[0028] S2.6 Update the sparrow position using firefly disturbance;

[0029] S2.7 calculates the fitness value and updates the sparrow position;

[0030] S2.8 determines whether the stop condition is met. If so, exit and output the result. Otherwise, repeat from S2.2.

[0031] Preferably, the variable mode decomposition in step S3 includes the following steps:

[0032] S3.1 Establish the constrained variation problem: Assume that the original signal f(t) is decomposed into K modal components {u} with different frequency characteristics. The sum of the estimated bandwidths of each mode must be minimized, while the sum of all modes must be equal to the original signal. Then the constrained variation expression is established as:

[0033]

[0034] S3.2 Solve the optimal solution of the constrained variational problem: First, perform Lagrange transformation to transform the above problem into an unconstrained variational problem for solution:

[0035]

[0036] In order to obtain the optimal solution to the problem, the alternating direction multiplier method is used to update each modal component and its center frequency. The formula is as follows:

[0037]

[0038] Where n is the number of iterations; τ is the update parameter, and the superscript ∧ represents the Fourier transform of the corresponding signal.

[0039] Preferably, the maximum information coefficient in step S5 includes the following contents:

[0040] S5.1 Assume that X and Y are two random variables in the data, where X = {X1,…,X n}, Y={Y1,…,Y n}, n is the number of samples; the mutual information between X and Y is defined as:

[0041]

[0042] Where p(x,y) is the joint probability density between X and Y; p(x) and p(y) represent the marginal probability densities of X and Y respectively;

[0043] S5.2 Draw a grid on the data scatter plot composed of variables X and Y, and calculate the mutual information between each grid; use different grid division standards to select the maximum value of the mutual information, and the calculation formula is:

[0044]

[0045] Among them, a and b represent the number of grids divided in the X and Y directions respectively, and B is the maximum value of the grid;

[0046] Preferably, the LSTM model and the Informer model in step S6 include the following contents:

[0047] S6.1 Forget gate uses the Sigmiod function to refer to the previous results and current content to decide whether the past state information needs to be discarded:

[0048] f t =σ(W m [h t-1 ,x t ]+b m )

[0049] Among them, x t is the information input at the current time t; h t-1 is the result of the last output; W m and b m are the forget gate weight matrix and bias term respectively; σ is the Sigmiod function; f tis the state of the forget gate. The closer its value is to 0, the more information is forgotten.

[0050] How much information is stored in the S6.2 memory cell is determined by the input gate:

[0051] i t =σ(W i [h t-1 ,x t ]+b i )

[0052]

[0053] Among them, W i , W c is the weight matrix; b i 、b c is the input gate bias term; tanh is the activation function; is a new memory unit; i t is the state of the input gate. The larger its value is, the more important the new information is, and unimportant information will be deleted.

[0054] S6.3 Before output is made, updated information will be determined, namely:

[0055]

[0056] Among them, c t For new memory units; is the current memory unit; the symbol “°” is the Hadamard multiplication;

[0057] S6.4 After confirming the information, the output gate will output to the external state:

[0058] o t =σ(W o [h t-1 ,x t ]+b o )

[0059] Among them, W o is the weight matrix; b o is the bias term; σ t is the state of the output gate. The closer the value is to 0, the greater the external state h t The more you can get from the memory unit c t Get more information in;

[0060] S6.5 Informer model proposes a new self-attention mechanism Prob Sparse self-attention, which is expressed as:

[0061]

[0062] in, and Q have the same size, that is, the computational complexity increases from O(L 2 ) is reduced to O(L ln L);

[0063] S6.6 In order to effectively save memory overhead and computing time, a distillation mechanism is introduced to gradually halve the sequence length. The expression of the distillation operation from the jth layer to the j+1th layer is:

[0064]

[0065] in, It includes multi-head sparse self-attention operations, Convld represents a one-dimensional convolution operation, ELU is the activation function, and MaxPool is the maximum pooling operation.

[0066] Preferably, the original data in step S1 is divided into two parts in a ratio of 9:1 after preprocessing, 90% of the original data after division in step S1 is used for prediction training of the model, and 10% of the original data after division in step S1 is used for detection and verification of the model.

[0067] Preferably, the original data is divided proportionally starting from any position, and the original data is divided proportionally for multiple times.

[0068] The beneficial effects of the present invention are as follows:

[0069] 1. The short-term power load forecasting method based on a hybrid model described in the present invention is based on a hybrid model of feature screening and frequency division prediction. The firefly sparrow algorithm is used to optimize the combination parameters in the variational mode decomposition to obtain the best decomposition effect signal component, thereby avoiding the influence of human experience, further improving the prediction accuracy, and substituting the best combination parameters into the variational mode decomposition to decompose the original power load data.

[0070] 2. The short-term power load forecasting method based on a hybrid model described in the present invention selects the separation points of high- and low-frequency signal components through an adaptive algorithm, screens the relevant characteristics of each signal component by the maximum information coefficient, obtains the input matrix of each signal component, and uses different models for forecasting, which not only reduces a certain amount of workload and reduces the forecasting time, but also can effectively improve the forecasting accuracy and overcome the problem of low prediction accuracy of a single model.

[0071] 3. The short-term power load forecasting method based on the hybrid model described in the present invention takes into account eight meteorological factors related to electricity price, temperature, humidity, etc. The rich characteristic information is conducive to reducing the prediction error and improving the universality of the overall model. BRIEF DESCRIPTION OF THE DRAWINGS

[0072] The present invention will be further described below in conjunction with the accompanying drawings.

[0073] Figure 1 It is a flow chart of the firefly sparrow algorithm to optimize the variational mode decomposition parameters;

[0074] Figure 2 It is the basic unit diagram of LSTM;

[0075] Figure 3 It is the Informer structure diagram;

[0076] Figure 4 It is the overall flow chart of the model;

[0077] Figure 5 It is a flow chart of the steps of the prediction method. DETAILED DESCRIPTION

[0078] In order to make the technical means, creative features, objectives and effects achieved by the present invention easy to understand, the present invention is further explained below in conjunction with specific implementation methods.

[0079] like Figures 1 to 5 As shown, the present invention provides a short-term power load forecasting method based on a hybrid model, and the forecasting method comprises the following steps:

[0080] S1: Preprocessing the original data; wherein the original data includes but is not limited to original power load data, real-time electricity price and weather factors, and the weather factors include but are not limited to temperature, humidity, rainfall, wind speed, maximum temperature, minimum temperature, average temperature and season;

[0081] The raw data are classified according to different administrative divisions, including provincial, municipal, county and township levels. After the raw data are classified, they are input into the model for prediction and multiple groups of prediction results are given;

[0082] By classifying the original data according to administrative divisions, we can obtain the power load data of different regions, such as the entire province, cities, counties, and towns within the province. After that, the classified original data is put into the model for prediction, so as to obtain relatively accurate power load data of different regions, which is convenient for power personnel to adjust the power grid and accurately allocate the power load of the province and various regions within the province, so as to ensure sufficient power supply and avoid excessive power load and waste of power resources.

[0083] Among them, the raw data preprocessing includes the following:

[0084] S1.1 uses linear interpolation to process outliers and missing values ​​and correct load data;

[0085] S1.2 In order to avoid the different dimensions of each eigenvalue, which affects the prediction accuracy and speeds up the gradient descent, the data is normalized uniformly:

[0086]

[0087] Among them, G' is the normalized data value; G is the original data value; G max and G min are the maximum and minimum values ​​in the original data set;

[0088] S1.3 Observe the daily average load curve of the original data and determine whether it has seasonality, periodicity, etc.;

[0089] S2: Use the firefly sparrow algorithm to optimize the parameter combination [k, α] in the variational mode decomposition and determine the best parameter combination [k, α];

[0090] Among them, the firefly sparrow algorithm includes the following steps:

[0091] S2.1 Initialization parameters: set the population size, number of iterations and dimension, initialize the ratio of predators and joiners, calculate the fitness value, and sort;

[0092] S2.2 Sparrow updates predator location;

[0093] S2.3 Sparrow updates the joiner's position;

[0094] S2.4 Sparrow updates the position of the alerter;

[0095] S2.5 calculates the fitness value and updates the sparrow position;

[0096] S2.6 Update the sparrow position using firefly disturbance;

[0097] S2.7 calculates the fitness value and updates the sparrow position;

[0098] S2.8 determines whether the stop condition is met. If so, exit and output the result. Otherwise, repeat from S2.2.

[0099] Specifically, the sparrow finder location is updated as follows:

[0100] The update formula of the discoverer position is:

[0101]

[0102] in, represents the position of the i-th sparrow in the j-th dimension in the t-th generation, α is a random number in (0,1), iter maxrepresents the maximum number of iterations, R2 represents the alarm value, ST represents the safety value, Q is a random number that obeys the normal distribution, and L represents a 1×n all-1 matrix;

[0103] The follower's position update formula is:

[0104]

[0105] in, represents the position of the individual with the worst fitness in the tth generation, represents the position of the individual with the best fitness in the t+1th generation, A represents a 1×dim matrix, in which the value of each element is randomly -1 or 1, A + =A T (AA T ) -1 ;

[0106] The position update formula of the sentinel is:

[0107]

[0108] in, represents the global optimal position in the tth generation, β is the control step size, which is a random number that obeys a normal distribution with a mean of 0 and a variance of 1, k∈[-1,1], ε is a constant to avoid the denominator being 0, f i represents the fitness value of the current individual, f g represents the fitness value of the current global optimal individual, f w Indicates the fitness value of the worst individual in the world at present;

[0109] Specifically, the relative fluorescence brightness of the firefly movement is:

[0110]

[0111] Where I0 is the maximum brightness of fireflies; γ is the light absorption coefficient; r i,j is the spatial distance between fireflies i and j;

[0112] The attractiveness of fireflies is:

[0113]

[0114] Among them, β0 is the maximum attraction of fireflies;

[0115] The motion equation of firefly i being attracted to firefly j is:

[0116]

[0117] Among them, x i , x jis the spatial position of fireflies i and j; α is the step factor, a random number in [0,1]; rand is a random number uniformly distributed on [0,1];

[0118] S3: Decomposing the preprocessed raw power load data using variational mode decomposition;

[0119] The variable mode decomposition includes the following steps:

[0120] S3.1 Establish the constrained variation problem: Assume that the original signal f(t) is decomposed into K modal components {u} with different frequency characteristics. The sum of the estimated bandwidths of each mode must be minimized, while the sum of all modes must be equal to the original signal. Then the constrained variation expression is established as:

[0121]

[0122] Among them, {u k} is the kth modal component after signal decomposition; {w k} is the frequency center of the kth modal component; δ(t) represents the Dirac distribution; * is the convolution operator;

[0123] S3.2 Solve the optimal solution of the constrained variational problem: First, perform Lagrange transformation to transform the above problem into an unconstrained variational problem for solution:

[0124]

[0125] In order to obtain the optimal solution to the problem, the alternating direction multiplier method is used to update each modal component and its center frequency. The formula is as follows:

[0126]

[0127] Where n is the number of iterations; τ is the update parameter, and the superscript ∧ represents the Fourier transform of the corresponding signal;

[0128] S4: using an adaptive algorithm to determine the dividing point between high-frequency and low-frequency signal components, and dividing the decomposed power load data into a high-frequency signal component and a low-frequency signal component according to the frequency dividing point;

[0129] S5: using the maximum information coefficient to select 3 to 5 features with the highest correlation for each signal component from among the relevant features, and establishing an input matrix;

[0130] Among them, the maximum information coefficient includes the following:

[0131] S5.1 Assume that X and Y are two random variables in the data, where X = {X1,…,X n}, Y={Y1,…,Y n}, n is the number of samples; the mutual information between X and Y is defined as:

[0132]

[0133] Where p(x,y) is the joint probability density between X and Y; p(x) and p(y) represent the marginal probability densities of X and Y respectively;

[0134] S5.2 Draw a grid on the data scatter plot composed of variables X and Y, and calculate the mutual information between each grid; use different grid division standards to select the maximum value of the mutual information, and the calculation formula is:

[0135]

[0136] Among them, a and b represent the number of grids divided in the X and Y directions respectively, and B is the maximum value of the grid;

[0137] S6: Input the low-frequency signal input matrix into the LSTM model for prediction, and output the prediction result of the LSTM model on the low-frequency signal input matrix; input the high-frequency signal input matrix into the Informer model for prediction, and output the prediction result of the Informer model on the high-frequency signal input matrix;

[0138] Among them, the LSTM model and Informer model include the following:

[0139] S6.1 Forget gate uses the Sigmiod function to refer to the previous results and current content to decide whether the past state information needs to be discarded:

[0140] f t =σ(W m [h t-1 ,x t ]+b m )

[0141] Among them, x t is the information input at the current time t; h t-1 is the result of the last output; W m and b m are the forget gate weight matrix and bias term respectively; σ is the Sigmiod function; f t is the state of the forget gate. The closer its value is to 0, the more information is forgotten.

[0142] How much information is stored in the S6.2 memory cell is determined by the input gate:

[0143] i t =σ(W i [h t-1 ,x t ]+bi )

[0144]

[0145] Among them, W i , W c is the weight matrix; b i 、b c is the input gate bias term; tanh is the activation function; is a new memory unit; i t is the state of the input gate. The larger its value is, the more important the new information is, and unimportant information will be deleted.

[0146] S6.3 Before output is made, updated information will be determined, namely:

[0147]

[0148] Among them, c t For new memory units; is the current memory unit; the symbol “°” is the Hadamard multiplication;

[0149] S6.4 After confirming the information, the output gate will output to the external state:

[0150] o t =σ(W o [h t-1 ,x t ]+b o )

[0151] Among them, W o is the weight matrix; b o is the bias term; σ t is the state of the output gate. The closer the value is to 0, the greater the external state h t The more you can get from the memory unit c t Get more information in;

[0152] S6.5 Informer model proposes a new self-attention mechanism Prob Sparse self-attention, which is expressed as:

[0153]

[0154] in, and Q have the same size, that is, the computational complexity increases from O(L 2 ) is reduced to O(LlnL);

[0155] S6.6 In order to effectively save memory overhead and computing time, a distillation mechanism is introduced to gradually halve the sequence length. The expression of the distillation operation from the jth layer to the j+1th layer is:

[0156]

[0157] in, It includes multi-head sparse self-attention operations, Convld represents a one-dimensional convolution operation, ELU is the activation function, and MaxPool is the maximum pooling operation;

[0158] S7: The prediction results of the LSTM model for the low-frequency signal input matrix and the prediction results of the Informer model for the high-frequency signal input matrix are superimposed and reorganized to output the final power load prediction results.

[0159] As an implementation mode of the present invention, the raw data in step S1 is divided into two parts according to a ratio of 9:1 after preprocessing, 90% of the raw data after division in step S1 is used for prediction training of the model, and 10% of the raw data after division in step S1 is used for detection and verification of the model;

[0160] By dividing the original data into two parts, the model is trained with the majority of the divided original data to obtain a prediction model with good performance. The obtained prediction model is then verified with the minority of the divided original data to observe the deviation between the prediction model and the actual situation. The relevant parameters in the prediction model are then corrected according to the deviation, thereby improving the fit between the prediction model and the actual situation and improving the accuracy of short-term power load prediction.

[0161] As an implementation mode of the present invention, the original data is divided proportionally starting from any position, and the original data is divided proportionally for multiple times;

[0162] Starting from any position in the preprocessed original data, division is performed in a ratio of 9:1, so that there is a difference between the data used for prediction training and the data for detection and verification after multiple divisions. Furthermore, after the model is trained with multiple groups of data used for prediction training obtained after division, the quality of the prediction model can be improved. When the prediction model is tested and verified with multiple groups of data used for detection and verification, the prediction model can be tested from multiple angles to find the shortcomings of the prediction model, and the relevant parameters in the prediction model can be adjusted as much as possible to make the prediction model fit the actual situation, thereby further improving the accuracy of short-term power load prediction.

[0163] The above shows and describes the basic principles, main features and advantages of the present invention. It should be understood by those skilled in the art that the present invention is not limited to the above embodiments. The above embodiments and descriptions are only for explaining the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements, which fall within the scope of the present invention. The scope of protection of the present invention is defined by the attached claims and their equivalents.

Claims

1. A short-term power load forecasting method based on a hybrid model, characterized in that: The prediction method comprises the following steps: S1: preprocess the raw data; S2: Use the firefly sparrow algorithm to optimize the parameter combination [k, α] in the variational mode decomposition and determine the best parameter combination [k, α]; S3: Decomposing the preprocessed raw power load data using variational mode decomposition; S4: using an adaptive algorithm to determine the dividing point between high-frequency and low-frequency signal components, and dividing the decomposed power load data into a high-frequency signal component and a low-frequency signal component according to the frequency dividing point; S5: using the maximum information coefficient to select 3 to 5 features with the highest correlation for each signal component from among the relevant features, and establishing an input matrix; S6: Input the low-frequency signal input matrix into the LSTM model for prediction, and output the prediction result of the LSTM model on the low-frequency signal input matrix; input the high-frequency signal input matrix into the Informer model for prediction, and output the prediction result of the Informer model on the high-frequency signal input matrix; S7: The prediction results of the LSTM model on the low-frequency signal input matrix and the prediction results of the Informer model on the high-frequency signal input matrix are superimposed and reorganized to output the final power load prediction result; The original data includes but is not limited to original power load data, real-time electricity prices and weather factors, and the weather factors include but are not limited to temperature, humidity, rainfall, wind speed, maximum temperature, minimum temperature, average temperature, and season; The original data are classified according to different administrative divisions. After the classification, the original data are respectively input into the model for prediction and multiple groups of prediction results are given.

2. A short-term power load forecasting method based on a hybrid model according to claim 1, characterized in that: The raw data preprocessing in step S1 includes the following contents: S1.1 uses linear interpolation to process outliers and missing values ​​and correct load data; S1.2 In order to avoid the different dimensions of each eigenvalue, which affects the prediction accuracy and speeds up the gradient descent, the data is normalized uniformly: Among them, G' is the normalized data value; G is the original data value; G max and G min are the maximum and minimum values ​​in the original data set; S1.3 Observe the daily average load curve of the original data and determine whether it has seasonality, periodicity, etc.

3. The short-term power load forecasting method based on the hybrid model according to claim 1 is characterized by: The firefly sparrow algorithm in step S2 includes the following steps: S2.1 Initialization parameters: set the population size, number of iterations and dimension, initialize the ratio of predators and joiners, calculate the fitness value, and sort; S2.2 Sparrow updates predator location; S2.3 Sparrow updates the joiner's position; S2.4 Sparrow updates the position of the alerter; S2.5 calculates the fitness value and updates the sparrow position; S2.6 Update the sparrow position using firefly disturbance; S2.7 calculates the fitness value and updates the sparrow position; S2.8 determines whether the stop condition is met. If so, exit and output the result. Otherwise, repeat from S2.

2.

4. The short-term power load forecasting method based on the hybrid model according to claim 1 is characterized by: The variable mode decomposition in step S3 includes the following steps: S3.1 Establish the constrained variation problem: Assume that the original signal f(t) is decomposed into K modal components {u} with different frequency characteristics. The sum of the estimated bandwidths of each mode must be minimized, while the sum of all modes must be equal to the original signal. Then the constrained variation expression is established as: S3.2 Solve the optimal solution of the constrained variational problem: First, perform Lagrange transformation to transform the above problem into an unconstrained variational problem for solution: In order to obtain the optimal solution to the problem, the alternating direction multiplier method is used to update each modal component and its center frequency. The formula is as follows: Where n is the number of iterations; τ is the update parameter, and the superscript ∧ represents the Fourier transform of the corresponding signal.

5. The short-term power load forecasting method based on a hybrid model according to claim 1 is characterized by: The maximum information coefficient in step S5 includes the following: S5.1 Assume that X and Y are two random variables in the data, where X = {X1,…,X n }, Y={Y1,…,Y n }, n is the number of samples; the mutual information between X and Y is defined as: Where p(x,y) is the joint probability density between X and Y; p(x) and p(y) represent the marginal probability densities of X and Y respectively; S5.2 Draw a grid on the data scatter plot composed of variables X and Y, and calculate the mutual information between each grid; use different grid division standards to select the maximum value of the mutual information, and the calculation formula is: Where a and b represent the number of grids divided in the X and Y directions respectively, and B is the maximum value of the grid.

6. The short-term power load forecasting method based on a hybrid model according to claim 1 is characterized by: The LSTM model and the Informer model in step S6 include the following: S6.1 Forget gate uses the Sigmiod function to refer to the previous results and current content to decide whether the past state information needs to be discarded: f t =σ(W m [h t-1 ,x t ]+b m ) Among them, x t is the information input at the current time t; h t-1 is the result of the last output; W m and b m are the forget gate weight matrix and bias term respectively; σ is the Sigmiod function; f t is the state of the forget gate. The closer its value is to 0, the more information is forgotten. How much information is stored in the S6.2 memory cell is determined by the input gate: i t =σ(W i [h t-1 ,x t ]+b i ) Among them, W i , W c is the weight matrix; b i 、b c is the input gate bias term; tanh is the activation function; is a new memory unit; i t is the state of the input gate. The larger its value is, the more important the new information is, and unimportant information will be deleted. S6.3 Before output is made, updated information will be determined, namely: Among them, c t For new memory units; is the current memory unit; the symbol "°" is the Hadamard multiplication; S6.4 After confirming the information, the output gate will output to the external state: the t =σ(W o [h t-1 ,x t ]+b o ) Among them, W o is the weight matrix; b o is the bias term; σ t is the state of the output gate. The closer the value is to 0, the greater the external state h t The more you can get from the memory unit c t Get more information in; S6.5 Informer model proposes a new self-attention mechanism Prob Sparse self-attention, which is expressed as: in, and Q have the same size, that is, the computational complexity increases from O(L 2 ) is reduced to O(LlnL); S6.6 In order to effectively save memory overhead and computing time, a distillation mechanism is introduced to gradually halve the sequence length. The expression of the distillation operation from the jth layer to the j+1th layer is: in, It includes multi-head sparse self-attention operations, Convld represents a one-dimensional convolution operation, ELU is the activation function, and MaxPool is the maximum pooling operation.

7. The short-term power load forecasting method based on a hybrid model according to claim 1 is characterized by: The original data in step S1 is divided into two parts in a ratio of 9:1 after preprocessing, 90% of the original data after division in step S1 is used for prediction training of the model, and 10% of the original data after division in step S1 is used for detection and verification of the model.

8. The short-term power load forecasting method based on the hybrid model according to claim 7 is characterized by: The original data is divided proportionally starting from any position, and the original data is divided proportionally for multiple times.

Images