Sub-industry load prediction method and system based on time sequence decomposition and multi-core Transform fusion, storage medium and electronic equipment

Abstract

The invention discloses a sub-industry load prediction method based on time sequence decomposition and multi-core Transform fusion. Firstly, industry loads are divided into a new energy dominant type, an industrial dominant type, a meteorological sensitive type and a mixed type through a clustering algorithm in combination with correlation coefficients of new energy output, policy factors and meteorological elements; secondly, decomposing the load data into trend components and day, weekly and monthly periodic components by adopting a time sequence decomposition layer, and respectively extracting features through a multi-kernel Transform architecture: capturing short-term fluctuation features by a time convolution kernel, capturing long-term trend features by a Transform layer based on a multi-head self-attention mechanism, fusing and outputting sub-industry load prediction results, and finally, accumulating four types of industry load prediction results to obtain a system load prediction result. According to the method, the problems of periodic component aliasing, difficulty in multi-factor coupling classification, insufficient long-term and short-term feature fusion and the like of a traditional method are solved, and the precision and robustness of sub-industry load prediction are improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of power system load forecasting, and particularly relates to a sub-industry load forecasting method based on the fusion of time series decomposition and multi-core Transformers. Background Art

[0002] With the deepening of the energy structure transformation and the power market reform, the power system load forecasting faces new challenges. On the one hand, the proportion of new energy power generation is increasing continuously, and the electricity load structure is becoming increasingly complex. Traditional load forecasting methods are difficult to meet the forecasting requirements under the new situation. On the other hand, with the deepening of the power market reform, the accuracy requirements for sub-industry and sub-region load forecasting are constantly increasing.

[0003] Existing traditional clustering methods (such as K-means, hierarchical clustering) are difficult to effectively handle the load classification problem with multiple factor couplings, especially the sub-industry load classification under the combined influence of new energy factors, policy factors, and meteorological factors. And single-modal decomposition techniques (such as EMD, VMD) are vulnerable to noise interference when dealing with complex load time series data, resulting in the aliasing of periodic components and affecting the subsequent forecasting accuracy. Existing forecasting models (such as LSTM, GRU) are difficult to effectively capture both the short-term mutation characteristics and long-term trend characteristics of load data simultaneously, especially the ability to fuse multi-time scale characteristics is insufficient.

[0004] Therefore, there is an urgent need to develop a load forecasting method that can effectively process multi-time scale time domain characteristics, adapt to various load types, and has strong generalization ability to meet the actual needs of power system load forecasting under the new situation. Summary of the Invention

[0005] Object of the Invention: The present invention proposes a time series prediction neural network based on the fusion of time series decomposition and multi-core Transformers to solve the problems of periodic component aliasing and insufficient long-term and short-term feature fusion in traditional load forecasting methods.

[0006] Technical Solution:

[0007] In a first aspect, the present invention discloses a sub-industry load forecasting method based on the fusion of time series decomposition and multi-core Transformers, including:

[0008] Collect historical data of industry loads in different industries and classify the industry loads into different types;

[0009] Construct a neural network model for time series prediction, and the neural network model is obtained based on the fusion of time series decomposition and multi-core Transformers;

[0010] Construct corresponding load datasets for the different types of industrial loads respectively, and use the load datasets to train the neural network model;

[0011] Input different types of industrial load data into the trained neural network model to obtain the load prediction results for this industrial type, and superimpose the load prediction results of each industrial type to obtain the system total load prediction results.

[0012] Further, divide the industrial loads into different types based on the historical data, including the steps of:

[0013] (1) For the historical data of the load data of each industry, calculate the correlation coefficients with new energy output, policy factors, and meteorological elements respectively, and the correlation coefficients are calculated using the Pearson correlation coefficient formula;

[0014] (2) Construct the multi-dimensional correlation feature vectors of each industry, and construct the corresponding multi-dimensional correlation feature vectors for each industry according to the correlation coefficients of new energy output, policy factors, and meteorological elements, including:

[0015] The multi-dimensional correlation feature vector corresponding to the new energy coefficient of industrial load n is where Load n represents the load of the nth type of industry, and solar represents the distributed photovoltaic data;

[0016] The multi-dimensional correlation feature vector corresponding to the policy factor coefficient is where policyEncoding represents the policy factor coding data;

[0017] The multi-dimensional correlation feature vector corresponding to the meteorological coefficient is where temperature represents the temperature data;

[0018] (3) Set threshold rules, and divide the industrial loads into different types based on the multi-dimensional correlation feature vectors, and the threshold rules are:

[0019] When and , the load type is defined as new energy dominant;

[0020] When and , the load type is defined as industry dominant;

[0021] When and , the load type is defined as meteorologically sensitive;

[0022] When the correlation coefficients do not meet any of the above conditions, the load type is defined as mixed.

[0023] Furthermore, constructing corresponding load datasets for the different categories of industrial loads includes: each category is divided into a training set, a validation set, and a test set according to a certain ratio, and the training set, the validation set, and the test set are respectively used to train the neural network model.

[0024] Furthermore, the neural network model includes a multi-temporal decomposition layer, a convolutional neural network layer, an encoder layer based on a multi-head self-attention mechanism, and a linear layer; the multi-temporal decomposition layer is used to decompose the input data into a trend component, a daily cycle component, a weekly cycle component, and a monthly cycle component, the convolutional neural network layer is used to extract short-term features of the input data according to the daily cycle component, the encoder layer is used to extract long-term features of the input data according to the weekly cycle component and the monthly cycle component, and the linear layer is used to extract trend features according to the trend component. Inputting different types of industrial load data into the trained neural network model to obtain the load prediction result of this industrial type, including the steps:

[0025] Inputting the time series load data Load ∈ R N·T , where N is the historical load and the feature dimension data, and T is the time step. Using the Fourier series algorithm to decompose the load data into a trend component Trend, a daily cycle component D daily , a weekly cycle component D weekly , and a monthly cycle component D monthly ;

[0026] Inputting the original input data Load and the daily cycle component D daily (t) into the convolutional network layer respectively to extract short-term features;

[0027] Inputting the weekly cycle component D weekly and the monthly cycle component D monthly into the encoder layer to extract long-term features;

[0028] The trend component Trend is mapped to the feature space through the linear layer to obtain the trend feature F trend ;

[0029] Concatenating the short-term feature F short , the long-term feature F long , and the trend feature F trend by concat to obtain the concatenated feature F concat , and outputting the prediction result Load pred through two layers of fully connected networks and the Sigmoid activation function.

[0030] Furthermore, using the Fourier series algorithm to decompose the load data into a trend component Trend and a daily cycle component D daily , a weekly cycle component D weekly , a monthly cycle component Dmonthly , specifically including:

[0031] (1) Trend component:

[0032]

[0033] Among them, t is the current moment, and i is the index value between -k and k of the moving average method;

[0034] (2) Daily cycle component:

[0035]

[0036] (3) Weekly cycle component:

[0037]

[0038] (4) Monthly cycle component:

[0039]

[0040] Among them, K1, K2, and K3 are used to control the number of harmonics, and a k , b k are the coefficients of the Fourier series formula.

[0041] Furthermore, the least squares method is used to jointly optimize all coefficients The optimization objective is:

[0042]

[0043] Among them, θ is the parameter set of a k , b k , is the coefficient of the daily cycle component D daily (t), is the coefficient of the weekly cycle component D weekly (t), is the coefficient of the monthly cycle component D monthly (t), and N is the number of samples.

[0044] Furthermore, the convolutional network layer is composed of several one-dimensional convolutional layers. The size of the convolutional kernel of each convolutional neural layer is set, and batch normalization and ReLU activation functions are connected after each convolution to obtain short-term features; the encoder layer includes several multi-head self-attention layers, with several attention heads in each layer, and the hidden layer dimension is set to obtain long-term features.

[0045] In the second aspect, the present invention discloses an industry-specific load forecasting system based on the fusion of time series decomposition and multi-core Transformer, including:

[0046] A classification module for collecting historical data of industrial loads in different industries and classifying industrial loads into different types based on the historical data;

[0047] A neural network model for performing time series prediction of loads, which is obtained by fusing time series decomposition and multi-core Transformer;

[0048] A training module for constructing corresponding load data sets for different types of industrial loads respectively and training the neural network model using the load data sets;

[0049] A prediction module for inputting different types of industrial load data into the trained neural network model to obtain the load prediction result of the industrial type, and superimposing the load prediction results of each industrial type to obtain the total system load prediction result.

[0050] Furthermore, classifying industrial loads into different types based on the historical data includes the steps of:

[0051] (1) For the historical data of the load data of each industry, calculate its correlation coefficients with new energy output, policy factors, and meteorological elements respectively, and the correlation coefficients are calculated using the Pearson correlation coefficient formula;

[0052] (2) Construct a multi-dimensional correlation feature vector for each industry, and construct a corresponding multi-dimensional correlation feature vector for each industry according to the correlation coefficients of new energy output, policy factors, and meteorological elements, including:

[0053] The multi-dimensional correlation feature vector corresponding to the new energy coefficient of industrial load n is where Load n represents the n-type industrial load, and solar represents distributed photovoltaic data;

[0054] The multi-dimensional correlation feature vector corresponding to the policy factor coefficient is where policyEncoding represents policy factor encoding data;

[0055] The multi-dimensional correlation feature vector corresponding to the meteorological coefficient is where temperature represents temperature data;

[0056] (3) Set a threshold rule, and classify industrial loads into different types based on the multi-dimensional correlation feature vector, and the threshold rule is:

[0057] When and the load type is defined as new energy dominant;

[0058] When And When, the load type is defined as industry - dominated;

[0059] When And When, the load type is defined as weather - sensitive;

[0060] When the correlation coefficient does not meet any of the above conditions, the load type is defined as mixed type.

[0061] Furthermore, constructing corresponding load datasets for the above - mentioned different categories of industrial loads includes: each category is divided into a training set, a validation set, and a test set according to a certain proportion, and the training set, the validation set, and the test set are used to train the neural network model respectively.

[0062] Furthermore, the neural network model includes a multi - time - series decomposition layer, a convolutional neural network layer, an encoder layer based on a multi - head self - attention mechanism, and a linear layer; the multi - time - series decomposition layer is used to decompose the input data into a trend component, a daily cycle component, a weekly cycle component, and a monthly cycle component, the convolutional neural network layer is used to extract short - term features of the input data according to the daily cycle component, the encoder layer is used to extract long - term features of the input data according to the weekly cycle component and the monthly cycle component, the linear layer is used to extract trend features according to the trend component. Inputting different types of industrial load data into the trained neural network model to obtain the load prediction result of this industry type, including the steps:

[0063] Input time - series load data Load ∈ R N·T , where N is the historical load and the feature dimension data, T is the time step, and the Fourier series algorithm is used to decompose the load data into a trend component Trend, a daily cycle component D daily , a weekly cycle component D weekly and a monthly cycle component D monthly ;

[0064] Input the original input data Load and the daily cycle component D daily (t) into the convolutional network layer respectively to extract short - term features;

[0065] Input the weekly cycle component D weekly and the monthly cycle component D monthly into the encoder layer to extract long - term features;

[0066] The trend component Trend is mapped to the feature space through the linear layer to obtain the trend feature F trend ;

[0067] Input the short - term feature F short , the long - term feature F long and the trend feature F trend Adopt the concat method to splice and obtain the spliced feature Fconcat , the prediction result is output through a two - layer fully - connected network and a Sigmoid activation function. Load pred .

[0068] Furthermore, the Fourier series algorithm is used to decompose the load data into a trend component Trend and daily - cycle components D daily , weekly - cycle components D weekly , and monthly - cycle components D monthly , specifically including:

[0069] (1) Trend component:

[0070]

[0071] where t is the current time, and i is the index value between - k and k in the moving average method;

[0072] (2) Daily - cycle component:

[0073]

[0074] (3) Weekly - cycle component:

[0075]

[0076] (4) Monthly - cycle component:

[0077]

[0078] where K1, K2, and K3 are used to control the number of harmonics, and a k , b k are the coefficients of the Fourier series formula.

[0079] Furthermore, the least - squares method is used to jointly optimize all coefficients The optimization objective is:

[0080]

[0081] where θ is the parameter set of a k , b k , is the coefficient of the daily - cycle component D daily (t), is the coefficient of the weekly - cycle component D weekly (t), is the coefficient of the monthly - cycle component D monthly (t), and N is the number of samples.

[0082] Further, the convolutional network layer is composed of several one-dimensional convolutional layers. The size of the convolutional kernel of each convolutional neural layer is set respectively. After each layer of convolution, batch normalization and ReLU activation function are connected to obtain short-term features. The encoder layer contains several multi-head self-attention layers, with several attention heads in each layer, and the hidden layer dimension is set to obtain long-term features.

[0083] In a third aspect, the present invention discloses a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the steps of the aforementioned load forecasting method for different industries are implemented.

[0084] In a fourth aspect, an electronic device is characterized by comprising: a processor and a memory. The memory stores one or more computer-executable instructions, and the processor calls the one or more computer-executable instructions to execute the steps of the aforementioned load forecasting method for different industries.

[0085] Beneficial effects:

[0086] According to the correlation coefficients of load with new energy output, policy factors and meteorological elements, the present invention divides the load into four categories and implements the strategy of "classified prediction - aggregation and superposition", scientifically classifying and aggregating the loads of different industries, and superimposing the prediction results of various loads to obtain the system total load prediction result, effectively alleviating the problem of multi-factor mixing of grid loads.

[0087] The time series prediction neural network based on time series decomposition and multi-core Transformer fusion proposed by the present invention captures the short-term fluctuation characteristics and long-term trend characteristics of load data simultaneously through time series decomposition technology and multi-core architecture, solves the problem that traditional prediction models cannot process multi-scale time series characteristics simultaneously, and significantly improves the prediction accuracy. Brief description of the drawings

[0088] Figure 1 It is a schematic flow chart of a load forecasting method for different industries based on time series decomposition and multi-core Transformer fusion of the present invention;

[0089] Figure 2 It is a schematic working diagram of the time series prediction neural network model of the present invention. Detailed implementation manners

[0090] The following further clarifies the present invention in conjunction with the drawings and specific implementation manners. It should be understood that the following specific implementation manners are only used to illustrate the present invention and not to limit the scope of the present invention. After reading the present invention, various equivalent forms of modification of the present invention by those skilled in the art all fall within the scope defined by the appended claims of this application.

[0091] The present invention discloses a load forecasting method for different industries based on the fusion of time series decomposition and multi-core Transformers, and its process is as follows: Figure 1 as shown, including:

[0092] Collect historical data of industry loads in different industries and divide the industry loads into different types;

[0093] Construct a neural network model for time series forecasting, including a multi-time series decomposition layer, a convolutional neural network layer, an encoder layer based on the multi-head self-attention mechanism, and a linear layer; the multi-time series decomposition layer is used to decompose the input data into a trend component, a daily cycle component, a weekly cycle component, and a monthly cycle component, the convolutional neural network layer is used to extract short-term features of the input data according to the daily cycle component, the encoder layer is used to extract long-term features of the input data according to the weekly cycle component and the monthly cycle component, and the linear layer is used to extract trend features according to the trend component;

[0094] Construct corresponding load data sets for the different types of industry loads respectively, and use the load data sets to train the neural network model;

[0095] Input different types of industry load data into the trained neural network model to obtain the load forecasting results of this industry type, and superimpose the load forecasting results of each industry type to obtain the total system load forecasting result.

[0096] Next, the technical solution of the present invention will be further described in combination with a specific embodiment and specific implementation steps. The steps of this embodiment are as follows:

[0097] S1 Collect historical load data of different industries and related features such as temperature, humidity, and irradiance.

[0098] S2 Perform industry clustering based on the characteristic analysis of industry load data, and divide the industry loads into four categories: new energy-dominated, industry-dominated, meteorologically sensitive, and mixed. The specific classification method is as follows:

[0099] (1) For the historical data of load data in different industries, calculate the correlation coefficients between the load data of each industry and new energy output, policy factors, and meteorological elements respectively, and construct a multi-dimensional correlation feature vector;

[0100] The correlation coefficient is calculated using the Pearson correlation coefficient formula:

[0101]

[0102] The multi-dimensional correlation feature vector includes:

[0103] The feature vector corresponding to the new energy coefficient of industry load n where Load nIndicates the load of n types of industries, and solar represents distributed photovoltaic data;

[0104] The eigenvector corresponding to the policy factors of the industry load n where policyEncoding represents the encoded data of policy factors;

[0105] The eigenvector corresponding to the meteorological coefficient of the industry load n where temperature represents temperature data.

[0106] (2) Based on the calculated correlation eigenvectors, according to the preset threshold rules, the industry load types are divided into new - energy - dominated, industry - dominated, meteorologically - sensitive, and mixed types. The specific preset threshold rules are shown in Table 1 as follows:

[0107] Table 1

[0108]

[0109] (3) Accumulate the industry load data of the same type to obtain the industry load data of each category.

[0110] S3 Construct a time - series prediction neural network model based on the fusion of time - series decomposition and multi - core Transformer.

[0111] This neural network model includes a multi - time - series decomposition layer, a convolutional neural network layer, an encoder layer based on the multi - head self - attention mechanism, and a linear layer, which is used to capture the dependency relationships of different cycles of time - series data in different regions.

[0112] S4 Construct load data sets for the four types of load data respectively, and divide them into training sets, validation sets, and test sets according to the ratio of 7:2:1.

[0113] S5 Use the load data sets to train the neural network model.

[0114] S6 Use the trained neural network model to perform time - series prediction on the load data of different types of industries. As Figure 2 shown is the working schematic diagram of the neural network model. In this embodiment, the specific working process of the neural network model is as follows:

[0115] (1) Multi - time - series decomposition layer: Select the load data of one type of industry from the load data set as the input load data Load∈R N·T , where N is the historical load and the feature dimension data, and T is the time step. Use the Fourier series algorithm to decompose the load data into a trend component Trend, a daily - cycle component D daily , a weekly - cycle component D weekly and a monthly - cycle component D monthly, the specific calculation method is as follows:

[0116] (1.1) Trend component Trend extraction. Calculate the trend component using the moving average method:

[0117]

[0118] Where t is the current time, and i is the index value between -k and k in the moving average method. In this embodiment, k = 2880, covering a one-month period.

[0119] (1.2) Period component separation. Remove the trend component from the original data:

[0120] R(t) = Load(t) - trend(t)

[0121] (1.3) Daily period component D daily Extraction. Use the Fourier series formula:

[0122]

[0123] Where, a k , b k Are the coefficients of the Fourier series formula, and K is used to control the number of harmonics, usually taking values between 3 and 5. In this embodiment, K1 = 5 in the daily period component extraction formula.

[0124] (1.4) Weekly period component D weekly Extraction. Use the Fourier series formula:

[0125]

[0126] Where K2 = 3.

[0127] (1.5) Monthly period component D monthly Extraction. Use the Fourier series formula:

[0128]

[0129] Where K3 = 3.

[0130] (1.6) Parameter optimization. Use the least squares method to jointly optimize the coefficients involved in steps (1.1) to (1.5), Establish an optimization objective using the square loss function, and the formula is expressed as:

[0131]

[0132] Where, θ is the set of ak and bk parameters in claim 5, Is the coefficient of the daily period component D daily (t), is the coefficient of the weekly periodic component D weekly (t), is the coefficient of the monthly periodic component D monthly (t), and N is the number of samples.

[0133] (1.7) uses linear filling to clean and correct outliers and missing values.

[0134] (2) Convolutional neural network layer: The original input data Load of this industry type and the daily periodic component D daily (t) are respectively input into the convolutional network layer to extract short-term features. In this embodiment, the convolutional network consists of 3 one-dimensional convolutional layers, the convolutional kernel sizes are 3, 5, and 7 respectively, and each layer of convolution is followed by batch normalization and ReLU activation function to obtain short-term feature F short .

[0135] (3) Multi-head self-attention encoder layer: The weekly periodic component D weekly and the monthly periodic component D monthly are input into the Transformer encoder to extract long-term features. In this embodiment, the encoder contains 4 multi-head self-attention layers, each layer has 8 attention heads, and the hidden layer dimension is 256, to obtain long-term feature F long .

[0136] (4) Linear layer: The trend component Trend is mapped to the feature space through a linear projection layer to obtain the trend feature F trend .

[0137] (5) Feature fusion and output: The short-term feature F short , the long-term feature F long and the trend feature F trend are concatenated by concat to obtain the concatenated feature F concat , and the load prediction result Load of this industry type is output through two fully connected networks and Sigmoid activation function pred .

[0138] After obtaining the load prediction results of four types, namely new energy-dominated type, industry-dominated type, temperature-dominated type, and mixed factor type, by S7 respectively, the predicted results are superimposed to obtain the system total load prediction result.

[0139] Through a clustering algorithm, the present invention combines the correlation coefficients of new energy output, policy factors, and meteorological elements to divide industrial loads into four categories: new energy-dominated type, industry-dominated type, meteorologically sensitive type, and mixed type. Secondly, a time series decomposition layer is used to decompose the load data into a trend component and daily, weekly, and monthly cycle components, and a multi-core Transformer architecture is used to extract features respectively: a temporal convolutional kernel captures short-term fluctuation features, and a Transformer layer based on a multi-head self-attention mechanism captures long-term trend features. The fused output obtains the load prediction results for different industries, and finally the load prediction results for the four types of industries are accumulated to obtain the total system load prediction result. The present invention solves problems such as aliasing of cycle components in traditional methods, difficulty in multi-factor coupled classification, and insufficient long-term and short-term feature fusion, and improves the accuracy and robustness of load prediction for different industries.

[0140] In a second aspect, the present invention also discloses a system for predicting loads for different industries based on the fusion of time series decomposition and multi-core Transformer, including:

[0141] A classification module for collecting historical data of industrial loads in different industries and dividing the industrial loads into different types based on the historical data;

[0142] A neural network model for performing time series prediction of loads, where the neural network model is obtained based on the fusion of time series decomposition and multi-core Transformer;

[0143] A training module for constructing corresponding load data sets for the industrial loads of different types respectively and training the neural network model using the load data sets;

[0144] A prediction module for inputting industrial load data of different types into the trained neural network model to obtain the load prediction result of the industrial type, and superimposing the load prediction results of each industrial type to obtain the total system load prediction result.

[0145] In a third aspect, the present invention also discloses a computer-readable storage medium on which a computer program is stored, and when the computer program is executed by a processor, the steps of the foregoing method for predicting loads for different industries are implemented.

[0146] In a fourth aspect, the present invention also discloses an electronic device, including: a processor and a memory, where the memory stores one or more computer-executable instructions, and the processor invokes the one or more computer-executable instructions to execute the steps of the foregoing method for predicting loads for different industries.

Claims

1. A load forecasting method for different industries based on the fusion of time series decomposition and multi-core Transformer, characterized in that Including: Collect historical data of industrial loads in different industries, and classify industrial loads into different types based on the historical data; Construct a neural network model for time series prediction, which is obtained by fusing time series decomposition and multi-core Transformer; Construct corresponding load data sets for different types of industrial loads respectively, and use the load data sets to train the neural network model; Input different types of industrial load data into the trained neural network model to obtain the load prediction results of the industrial type, and superimpose the load prediction results of each industrial type to obtain the total system load prediction results.

2. The load prediction method according to claim 1, characterized in that Classifying industrial loads into different types based on the historical data includes the steps of: (1) For the historical data of the load data of each industry, calculate its correlation coefficients with new energy output, policy factors, and meteorological elements respectively, and the correlation coefficients are calculated using the Pearson correlation coefficient formula; (2) Construct a multi-dimensional correlation feature vector for each industry, and construct a corresponding multi-dimensional correlation feature vector for each industry according to the correlation coefficients of new energy output, policy factors, and meteorological elements, including: The multi-dimensional correlation feature vector corresponding to the new energy coefficient of the industry load n is where Load n represents the n types of industry loads, and solar represents the distributed photovoltaic data; The multi-dimensional correlation eigenvector corresponding to the policy factor coefficient is where policyEncoding represents the policy factor coding data; The multi-dimensional correlation feature vector corresponding to the meteorological coefficient is where temperature represents temperature data; (3) Set threshold rules, and classify industrial loads into different types based on the multi-dimensional correlation feature vector, and the threshold rules are: When and the industry load type is defined as new energy - dominated; When and the industry load type is defined as industry-dominated; When and the industry load type is defined as meteorologically sensitive; When the correlation coefficient does not meet any of the above conditions, the load type is defined as a mixed type.

3. The load prediction method according to claim 2, wherein Constructing corresponding load data sets for different categories of industrial loads includes: each category is divided into a training set, a validation set, and a test set according to a certain proportion, and the training set, the validation set, and the test set are used to train the neural network model respectively.

4. The load prediction method according to claim 3, characterized in that, The neural network model includes a multi-time series decomposition layer, a convolutional neural network layer, an encoder layer based on a multi-head self-attention mechanism, and a linear layer; the multi-time series decomposition layer is used to decompose the input data into a trend component, a daily cycle component, a weekly cycle component, and a monthly cycle component, the convolutional neural network layer is used to extract short-term features of the input data according to the daily cycle component, the encoder layer is used to extract long-term features of the input data according to the weekly cycle component and the monthly cycle component, the linear layer is used to extract trend features according to the trend component, and input different types of industrial load data into the trained neural network model to obtain the load prediction results of the industrial type, including the steps of: Input time series load data Load ∈ R N·T , where N is the historical load and feature dimension data, T is the time step, and the Fourier series algorithm is used to decompose the load data into a trend component Trend, a daily cycle component D daily , a weekly cycle component D weekly and a monthly cycle component D monthly ; Load the original input data and the daily cycle component D daily Input them into the convolutional network layer respectively to extract short-term features; Input the weekly periodic component D weekly and the monthly periodic component D monthly into the encoder layer to extract long-term features; The Trend component is mapped to the feature space through a linear layer to obtain the trend feature F trend ; The short-term feature F short , the long-term feature F long and the trend feature F trend are concatenated by the concat method to obtain the concatenated feature F concat , and the load prediction result Load is output through a two-layer fully connected network and a Sigmoid activation function pred .

5. The load prediction method according to claim 4, characterized in that The load data is decomposed into a trend component Trend, a daily cycle component D daily , a weekly cycle component D weekly and a monthly cycle component D monthly by using the Fourier series algorithm, specifically including: (1) Trend component: where t is the current time and i is the index value between moving average method -k and k; (2) Daily cycle component: (3) Weekly cycle component: (4) Monthly cycle component: Among them, K1, K2, and K3 are used to control the number of harmonics, and a k , b k are the coefficients of the Fourier series formula.

6. The load prediction method according to claim 5, wherein Jointly optimize all coefficients using the least squares method The optimization objective is: where θ is a k , b k parameter set, is the coefficient of the daily cycle component D daily (t), is the coefficient of the weekly cycle component D weekly (t), is the coefficient of the monthly cycle component D monthly (t), and N is the number of samples.

7. The load prediction method according to claim 6, wherein The convolutional neural network layer is composed of several one-dimensional convolutional layers, and the size of the convolutional kernel of each convolutional layer is set respectively. After each layer of convolution, batch normalization and ReLU activation function are connected to obtain short-term features; The encoder layer contains several multi-head self-attention layers, several attention heads in each layer, and the hidden layer dimension is set to obtain long-term features.

8. An industry-specific load forecasting system based on time series decomposition and multi-core Transformer fusion, characterized in that, Including: A classification module for collecting historical data of industrial loads in different industries and classifying industrial loads into different types based on the historical data; A neural network model for performing time series prediction of load, which is obtained by fusing time series decomposition and multi-core Transformer; A training module for constructing corresponding load data sets for different types of industrial loads respectively, and using the load data sets to train the neural network model; A prediction module for inputting different types of industrial load data into the trained neural network model to obtain the load prediction results of the industrial type, and superimposing the load prediction results of each industrial type to obtain the total system load prediction result.

9. The load prediction method according to claim 8, wherein Dividing industrial loads into different types based on the historical data, including the steps of: (1) For the historical data of the load data of each industry, calculate its correlation coefficients with new energy output, policy factors and meteorological elements respectively, and the correlation coefficients are calculated using the Pearson correlation coefficient formula; (2) Construct a multi-dimensional correlation feature vector for each industry, and construct a corresponding multi-dimensional correlation feature vector for each industry according to the correlation coefficients of the new energy output, policy factors and meteorological elements, including: The multi-dimensional correlation feature vector corresponding to the new energy coefficient of the industry load n is where Load n represents the n types of industry loads, and solar represents the distributed photovoltaic data; The multi-dimensional correlation feature vector corresponding to the policy factor coefficient is where policyEncoding represents the policy factor encoding data; The multi-dimensional correlation feature vector corresponding to the meteorological coefficient is where temperature represents temperature data; (3) Set a threshold rule, and based on the multi-dimensional correlation feature vector, divide industrial loads into different types, and the threshold rule is: When and the load type is defined as new - energy - dominated; When and the load type is defined as industry-dominated; When and the load type is defined as meteorologically sensitive; When the correlation coefficient does not meet any of the above conditions, the load type is defined as a mixed type.

10. The load forecasting system according to claim 9, characterized in that, Constructing corresponding load data sets for different categories of industrial loads includes: each category is divided into a training set, a validation set and a test set according to a certain proportion, and the training set, the validation set and the test set are used to train the neural network model respectively.

11. The load forecasting system according to claim 10, wherein The neural network model includes a multi-time series decomposition layer, a convolutional neural network layer, an encoder layer based on a multi-head self-attention mechanism, and a linear layer; the multi-time series decomposition layer is used to decompose the input data into a trend component, a daily cycle component, a weekly cycle component and a monthly cycle component, the convolutional neural network layer is used to extract short-term features of the input data according to the daily cycle component, the encoder layer is used to extract long-term features of the input data according to the weekly cycle component and the monthly cycle component, and the linear layer is used to extract trend features according to the trend component. Inputting different types of industrial load data into the trained neural network model to obtain the load prediction results of the industrial type, including the steps of: Input time series load data Load ∈ R N·T , where N is the historical load and feature dimension data, T is the time step, and the Fourier series algorithm is used to decompose the load data into a trend component Trend, a daily cycle component D daily , a weekly cycle component D weekly , and a monthly cycle component D monthly ; Load the original input data and the daily cycle component D daily Input them into the convolutional network layer respectively to extract short-term features; Input the weekly periodic component D weekly and the monthly periodic component D monthly into the encoder layer to extract long-term features; The trend component Trend is mapped to the feature space through a linear layer to obtain the trend feature F trend ; Concatenate the short-term feature F short , the long-term feature F long and the trend feature F trend using the concat method to obtain the concatenated feature F concat , and output the load prediction result Load through a two-layer fully connected network and a Sigmoid activation function pred .

12. The load prediction system according to claim 11, wherein The Fourier series algorithm is used to decompose the load data into a trend component Trend and daily cycle components D daily , weekly cycle components D weekly , monthly cycle components D monthly , specifically including: (1) Trend component: where t is the current time, and i is the index value between moving average method -k and k; (2) Daily cycle component: (3) Weekly cycle component: (4) Monthly cycle component: Among them, K1, K2, and K3 are used to control the number of harmonics, and a k , b k are the coefficients of the Fourier series formula.

13. The load forecasting system according to claim 12, wherein Jointly optimize all coefficients using the least squares method The optimization objective is as follows: where θ is a k , b k parameter set, is the coefficient of the daily periodic component D daily (t), is the coefficient of the weekly periodic component D weekly (t), is the coefficient of the monthly periodic component D monthly (t), and N is the number of samples.

14. The load forecasting system according to claim 13, wherein The convolutional neural network layer is composed of several one-dimensional convolutional layers, and the size of the convolutional kernel of each convolutional layer is set respectively. After each convolution, batch normalization and ReLU activation function are connected to obtain short-term features; The encoder layer contains several multi-head self-attention layers, several attention heads in each layer, and the hidden layer dimension is set to obtain long-term features.

15. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method described in claims 1 to 7.

16. An electronic device, characterized in that, Including: A processor and a memory, the memory stores one or more computer-executable instructions, and the processor calls the one or more computer-executable instructions to execute the steps of the sub-industry load prediction method described in any one of claims 1 to 7.

Images