CN116865261B

CN116865261B - Power load prediction method and system based on twin network

Info

Publication number: CN116865261B
Application number: CN202310885152.3A
Authority: CN
Inventors: 王克佳; 李雪
Original assignee: Meizhou Jiaan Electric Power Design Co ltd
Current assignee: Meizhou Jiaan Electric Power Design Co ltd
Priority date: 2023-07-19
Filing date: 2023-07-19
Publication date: 2024-03-15
Anticipated expiration: 2043-07-19
Also published as: CN116865261A

Abstract

The invention provides a power load prediction method and system based on a twin network, and belongs to the technical field of power. Firstly, acquiring historical power load data and historical power satellite images; secondly, carrying out characteristic engineering treatment on the historical power load data to obtain time sequence data; extracting features through a satellite image network to obtain one-dimensional data, and combining the one-dimensional data with time sequence data to obtain power load sequence data; then inputting the training set into a twin network for power load prediction to obtain an initial power load prediction model; and finally, inputting the test set into the initial power load prediction model to test, and outputting the final power load prediction model until the set condition is met. By combining the satellite images and the time sequence data, the power load can be more comprehensively analyzed and predicted, the accuracy and the effect of the prediction model are improved, and accurate power demand estimation is provided for the operation and planning of a power system.

Description

Power load prediction method and system based on twin network

Technical Field

The invention belongs to the technical field of power, and particularly relates to a power load prediction method and system based on a twin network.

Background

Electrical load prediction is an estimate of electrical demand over a period of time in the future, which is critical to the operation and planning of an electrical power system; the accurate load prediction can help an electric power company to plan a reasonable power generation strategy, the electric power company needs to decide to start, stop or adjust the operation of a generator set according to the future load demand, and if the load prediction is inaccurate, excessive power generation or insufficient power generation can be caused, so that energy is wasted or the demand cannot be met; also facilitating optimization of power distribution and scheduling, in power systems, the goal of power distribution is to transfer power from a power plant to a consumer, while the goal of power scheduling is to ensure stable operation of the system. The accurate load prediction can help the power company reasonably arrange power distribution and scheduling strategies, and avoid the occurrence of the condition of overload or insufficient power supply.

In the prior art, chinese patent publication No.: CN114529051a provides a long-term power load prediction method based on hierarchical residual self-attention neural network. The patent includes: (1) the method comprises the steps of (1) adaptively extracting mixed characteristic data of trend items, period items, holiday items and weather items in historical load data, fusing the mixed characteristic data with the historical load sequence data, (2) carrying out time component recursion decomposition on the fused sequence data, encoding the time component by using a hierarchical residual self-attention network block, (3) reconstructing the time component, carrying out generative decoding, and predicting power load fluctuation of a period in the future.

However, this patent only analyzes and discusses the power load prediction with respect to historical load data and does not analyze the relevant power station image information, which may capture images of a panoramic view or a specific area of the power station from the air or space. The satellite images may provide spatial information of the power station and its surroundings, including layout and status of buildings, equipment, transmission lines, substations, etc. By analyzing and processing these images, by combining the power station satellite images and the time series data, the power load can be more comprehensively analyzed and predicted, and the accuracy and effect of the prediction model can be improved.

Disclosure of Invention

Based on the technical problems, the invention provides a power load prediction method and a power load prediction system based on a twin network, which are used for comprehensively analyzing and predicting power loads by combining historical power load data and power satellite image information, so that the prediction accuracy is improved.

The invention provides a twin network-based power load prediction method, which comprises the following steps:

step S1: acquiring historical power load data and historical power satellite images;

step S2: carrying out characteristic engineering processing on the historical power load data to obtain time sequence data;

Step S3: inputting the historical power satellite image into a satellite image network for feature extraction to obtain a feature image C28, and performing data conversion on the feature image to obtain one-dimensional image data;

step S4: performing data processing operation on the one-dimensional image data and the time sequence data to obtain power load sequence data;

step S5: respectively constructing a time sequence data set and a power load data set based on the time sequence data and the power load sequence data, wherein the time sequence data set is divided into a time sequence training set and a time sequence test set according to a certain proportion, and the power load data set is divided into a power load training set and a power load test set according to a certain proportion;

step S6: inputting the time sequence training set and the electric power load training set into a twin network to perform electric power load prediction to obtain an initial electric power load prediction model;

step S7: and inputting the time sequence test set and the power load test set into an initial power load prediction model for testing, and outputting a final power load prediction model until the set condition is met.

Optionally, the step of inputting the historical power satellite image into a satellite image network for feature extraction to obtain a feature map C28, and performing data conversion on the feature map to obtain one-dimensional image data specifically includes:

The satellite image network specifically comprises a first standard convolution module, a residual error module, an attention module and a second standard convolution module;

inputting the historical electric satellite image into the first standard convolution module for feature extraction to obtain a feature map C3;

inputting the characteristic diagram C3 into the residual error module for characteristic extraction to obtain a characteristic diagram C12;

inputting the feature map C12 into the attention module for feature extraction to obtain a feature map C25;

inputting the feature map C25 into the second standard convolution module to perform feature extraction to obtain a feature map C28;

and respectively carrying out numerical calculation on the characteristic map C28 to obtain a maximum value, a minimum value and an average value, wherein the one-dimensional image data comprises the maximum value, the minimum value and the average value.

Optionally, the inputting the feature map C12 to the attention module to perform feature extraction to obtain a feature map C25 specifically includes:

the attention module comprises an attention input layer, a vertical average pooling layer, a horizontal average pooling layer, a fifth standard convolution layer, a first dimension arrangement layer, a second dimension arrangement layer, a fifth normalized activation layer, a sixth standard convolution layer, a seventh standard convolution layer, a first activation function layer, a second activation function layer, a third dimension arrangement layer, a fourth dimension arrangement layer and an element multiplication layer;

Inputting the feature map C12 using the attention input layer;

inputting the feature map C12 to the vertical average pooling layer for vertical pooling to obtain a feature map C13, inputting the feature map C13 to the first dimension arrangement layer for dimension conversion to obtain a feature map C16, inputting the feature map C16 to the sixth standard convolution layer for convolution operation to obtain a feature map C19, inputting the feature map C19 to the first activation function layer for activation operation to obtain a feature map C21, and inputting the feature map C21 to the third dimension arrangement layer for dimension conversion to obtain a feature map C23;

inputting the feature map C12 to the horizontal average pooling layer for horizontal pooling to obtain a feature map C14, inputting the feature map C14 to the second dimension arrangement layer for dimension conversion to obtain a feature map C17, inputting the feature map C17 to the seventh standard convolution layer for convolution operation to obtain a feature map C20, inputting the feature map C20 to the second activation function layer for activation operation to obtain a feature map C22, and inputting the feature map C22 to the fourth dimension arrangement layer for dimension conversion to obtain a feature map C24;

Inputting the characteristic diagram C12 into the fifth standard convolution layer to carry out convolution operation to obtain a characteristic diagram C15, and inputting the characteristic diagram C15 into the fifth normalized activation layer to carry out batch normalization and activation operation to obtain a characteristic diagram C18;

and inputting the feature maps C23, C24 and C18 into the element multiplication layer to carry out element multiplication to obtain a feature map C25.

Optionally, the performing a data processing operation on the one-dimensional image data and the time series data to obtain power load sequence data specifically includes:

respectively carrying out data normalization processing on the maximum value, the minimum value and the average value in the one-dimensional image data to obtain a standard maximum value, a standard minimum value and a standard average value;

and carrying out data alignment on the standard maximum value, the standard minimum value and the standard average value and the time series data to obtain power load sequence data.

Optionally, the inputting the time sequence training set and the electric load training set into a twin network to perform electric load prediction to obtain an initial electric load prediction model, which specifically includes:

inputting the time sequence training set into a first branch of the twin network to perform feature extraction to obtain time sequence features;

Inputting the electric load training set into a second branch of the twin network for feature extraction to obtain electric load features;

performing feature fusion on the time sequence feature and the power load feature to obtain a power load fusion feature;

and predicting the power load fusion characteristics to obtain an initial power load prediction model.

Optionally, the inputting the time sequence training set to the first branch of the twin network performs feature extraction to obtain a time sequence feature, which specifically includes:

the first branch comprises a first residual error module and a second residual error module;

inputting the time sequence training set into the first residual error module to perform expansion causal convolution operation to obtain a feature map T1;

and inputting the characteristic diagram T1 into the second residual error module to perform expansion causal convolution operation, so as to obtain a time sequence characteristic.

The invention also provides a twin network-based power load prediction system, the system comprising:

the power data acquisition module is used for acquiring historical power load data and historical power satellite images;

the characteristic engineering processing module is used for carrying out characteristic engineering processing on the historical power load data to obtain time sequence data;

The power image module is used for inputting the historical power satellite image into a satellite image network for feature extraction to obtain a feature image C28, and performing data conversion on the feature image to obtain one-dimensional image data;

the data processing module is used for performing data processing operation on the one-dimensional image data and the time sequence data to obtain power load sequence data;

the data set construction module is used for respectively constructing a time sequence data set and a power load data set based on the time sequence data and the power load sequence data, wherein the time sequence data set is divided into a time sequence training set and a time sequence test set according to a certain proportion, and the power load data set is divided into a power load training set and a power load test set according to a certain proportion;

the initial prediction module is used for inputting the time sequence training set and the electric load training set into a twin network to perform electric load prediction to obtain an initial electric load prediction model;

and the test module is used for inputting the time sequence test set and the power load test set into an initial power load prediction model for testing, and outputting a final power load prediction model until the set condition is met.

Optionally, the power image module specifically includes:

the first standard convolution sub-module is used for inputting the historical electric satellite image into the first standard convolution sub-module for feature extraction to obtain a feature map C3;

the residual sub-module is used for inputting the characteristic diagram C3 into the residual module for characteristic extraction to obtain a characteristic diagram C12;

the attention sub-module is used for inputting the feature map C12 into the attention module for feature extraction to obtain a feature map C25;

the second standard convolution sub-module is used for inputting the feature map C25 into the second standard convolution module to perform feature extraction to obtain a feature map C28;

and the numerical calculation sub-module is used for respectively carrying out numerical calculation on the characteristic map C28 to obtain a maximum value, a minimum value and an average value, and the one-dimensional image data comprises the maximum value, the minimum value and the average value.

Optionally, the attention sub-module specifically includes:

an attention input unit that inputs the feature map C12 using the attention input layer;

the vertical average pooling unit is configured to input the feature map C12 to the vertical average pooling layer for vertical pooling to obtain a feature map C13, input the feature map C13 to the first dimension arrangement layer for dimension conversion to obtain a feature map C16, input the feature map C16 to a sixth standard convolution layer for convolution operation to obtain a feature map C19, input the feature map C19 to a first activation function layer for activation operation to obtain a feature map C21, and input the feature map C21 to a third dimension arrangement layer for dimension conversion to obtain a feature map C23;

The horizontal average pooling unit is configured to input the feature map C12 to the horizontal average pooling layer for horizontal pooling to obtain a feature map C14, input the feature map C14 to the second dimension arrangement layer for dimension conversion to obtain a feature map C17, input the feature map C17 to a seventh standard convolution layer for convolution operation to obtain a feature map C20, input the feature map C20 to a second activation function layer for activation operation to obtain a feature map C22, and input the feature map C22 to a fourth dimension arrangement layer for dimension conversion to obtain a feature map C24;

a fifth convolution normalized activation unit, configured to input the feature map C12 to the fifth standard convolution layer for convolution operation, obtain a feature map C15, and input the feature map C15 to the fifth normalized activation layer for batch normalization and activation operation, so as to obtain a feature map C18;

and the element multiplication unit is used for inputting the feature graphs C23, C24 and C18 into the element multiplication layer to carry out element multiplication to obtain a feature graph C25.

Optionally, the initial prediction module specifically includes:

the first branch submodule is used for inputting the time sequence training set into a first branch of the twin network to perform feature extraction so as to obtain time sequence features;

The second branch sub-module is used for inputting the electric load training set into a second branch of the twin network to perform feature extraction so as to obtain electric load features;

the characteristic fusion sub-module is used for carrying out characteristic fusion on the time sequence characteristic and the power load characteristic to obtain a power load fusion characteristic;

and the characteristic prediction sub-module is used for predicting the power load fusion characteristic to obtain an initial power load prediction model.

Compared with the prior art, the invention has the following beneficial effects:

the present invention uses a twin network, an expansion cause and effect convolution and an attention mechanism, where the twin network can be used to process historical power load data and image signature data. By inputting historical power load data and other relevant characteristic data into the two sub-networks, the twin network can learn the relevance and similarity between the two sub-networks, so that the prediction accuracy is improved; the dilation-causal convolution can capture long-term dependencies in the sequence data, the time sequence data has a time sequence characteristic, and the dilation-causal convolution can effectively process the time sequence. By increasing the receptive field of the convolution kernel, the expanding causal convolution can better capture the time sequence mode and trend in the load data under the condition of not increasing the number of parameters; the attention mechanism may help the network focus on features and time periods that have a greater contribution to the predicted load. By applying the attention mechanism to the historical satellite image feature data, the network can dynamically adjust the weight, so that more important data in the prediction can obtain higher attention, the accuracy of power load prediction is improved, and accurate power demand estimation is provided for the operation and planning of a power system.

Drawings

FIG. 1 is a flow chart of a twin network-based power load prediction method of the present invention;

FIG. 2 is a block diagram of a satellite image network of a twin network-based power load prediction method of the present invention;

FIG. 3 is a block diagram of an attention module of the twin network based power load prediction method of the present invention;

FIG. 4 is a diagram of a twin network architecture of a twin network-based power load prediction method of the present invention;

fig. 5 is a block diagram of a twin network-based power load prediction system of the present invention.

Description of the embodiments

The invention is further described below in connection with specific embodiments and the accompanying drawings, but the invention is not limited to these embodiments.

Examples

As shown in fig. 1, the invention discloses a power load prediction method based on a twin network, which comprises the following steps:

step S3: inputting the historical electric satellite image into a satellite image network for feature extraction to obtain a feature image C28, and performing data conversion on the feature image to obtain one-dimensional image data;

step S6: inputting the time sequence training set and the electric load training set into a twin network to perform electric load prediction to obtain an initial electric load prediction model;

step S7: and inputting the time sequence test set and the power load test set into an initial power load prediction model for testing, and outputting a final power load prediction model until the set conditions are met.

The steps are discussed in detail below:

step S1: acquiring historical power load data and historical power satellite images, wherein the historical power load data specifically comprises time stamp data, load values, average temperature, highest temperature, lowest temperature, wind speed, air pressure values, rainfall, snowfall, holidays and seasons; the historical power satellite images specifically include acquired satellite images.

Step S2: carrying out characteristic engineering processing on the historical power load data to obtain time sequence data, wherein the method specifically comprises the following steps of:

Step S21: carrying out normalization processing on the numerical weather original data and the load value, wherein the normalization formula is as follows:

in the method, in the process of the invention,is normalized data; />For the raw data (data to be normalized), +.>For the minimum value in the data, +.>For the maximum value in the data, the above numerical weather data includes an average temperature, a maximum temperature, a minimum temperature, a wind speed, an air pressure value, a rainfall amount, and a snowfall amount.

Step S22: the state-type raw data is subjected to one-hot encoding (one-hot), such as season and holiday, for example, is divided into spring, summer, autumn and winter, the corresponding codes are respectively [1, 0], [0,1, 0], [0,1, 0] and [0, 1].

Step S23: and according to the time stamp sequence, aligning each item of data of the processed load value, the average temperature, the highest temperature, the lowest temperature, the wind speed, the air pressure value, the rainfall, the snowfall, the seasons and the holidays with the corresponding time.

Step S3: and inputting the historical electric satellite image into a satellite image network for feature extraction to obtain a feature image C28, and performing data conversion on the feature image to obtain one-dimensional image data.

In fig. 2-3, conv2D represents a standard convolutional layer; strides represents the step size; the normalized activation layer comprises a batch normalization layer (Batch Normalization) and an activation function layer (Activation (Relu)), and the normalized activation layer selects a Relu activation function; an individual Activation function layer (action ] ))，/>The values are Relu and Sigmoid; sepConv2D represents a depth separable convolutional layer; flat stands for flattening layer; dense stands for full connectivity layer; />AveragePooling2D，/>The value V or H respectively represents a vertical average pooling layer and a horizontal average pooling layer; concat (+)>,/>) Representation->And->Splicing; multiple (>,/>,/>) Representation->，/>，/>Performing element-by-element multiplication; maxpooling2D represents the maximum pooling layer; c->Representing the feature maps obtained in the satellite image network, respectively>The value range is [1,28 ]]，/>Is an integer; permute (C->:/>,,/>) Representing the characteristic diagram C->Go->,/>、/>,/>And->，/>Transformation of any one dimension.

As shown in fig. 2, step S3 specifically includes:

step S31: the historical electric satellite image is input into a first standard convolution module to perform feature extraction to obtain a feature map C3, and the method specifically comprises the following steps:

inputting a historical electric satellite image (256,256,3) into a first standard convolution layer for convolution operation to obtain a feature map C1, wherein the number of convolution kernels of the first standard convolution layer is 32, the size of the convolution kernels is 3 multiplied by 3, and the step length is 2; feature map C1 is 128×128 for 32 channels; inputting the feature map C1 into a first normalized activation layer to perform batch normalization and activation operation to obtain a feature map C2; feature map C2 is 128×128 for 32 channels; inputting the feature map C2 into a first maximum pooling layer to perform maximum pooling operation to obtain a feature map C3; the size of the pooling window of the first maximum pooling layer is 3 multiplied by 3, and the step length is 2; the feature map C3 is 64×64 of 32 channels.

In this embodiment, the first standard convolution module includes a first standard convolution layer, a first normalized activation layer, and a first max-pooling layer.

Step S32: inputting the feature map C3 into a residual error module for feature extraction to obtain a feature map C12, wherein the feature map C12 specifically comprises:

i, inputting the characteristic diagram C3 into a second standard convolution layer to carry out convolution operation to obtain a characteristic diagram C4, wherein the number of convolution kernels of the second standard convolution layer is 16, the size of the convolution kernels is 1 multiplied by 1, and the step length is 1; feature map C4 is 64×64 for 16 channels; inputting the feature map C4 into a second normalized activation layer to perform batch normalization and activation operation to obtain a feature map C5; feature map C5 is 64×64 for 16 channels; sequentially inputting the feature map C5 into a first depth separable convolution layer and a first batch of normalization layers to perform depth separable convolution operation and batch normalization operation to obtain a feature map C6; the number of convolution kernels of the first depth separable convolution layer is 16, the size of the convolution kernels is 3 multiplied by 3, and the step length is 2; feature map C6 is 32×32 for 16 channels; inputting the characteristic diagram C6 into a third standard convolution layer to carry out convolution operation to obtain a characteristic diagram C7, wherein the number of convolution kernels of the third standard convolution layer is 32, the size of the convolution kernels is 1 multiplied by 1, and the step length is 1; feature map C7 is 32×32 for 32 channels; inputting the feature map C7 into a third normalized activation layer to perform batch normalization and activation operation to obtain a feature map C8; the feature map C8 is 32×32 of 32 channels.

II, sequentially inputting the feature map C3 into a second depth separable convolution layer and a second batch of normalization layers to perform depth separable convolution operation and batch normalization operation to obtain a feature map C9; the number of convolution kernels of the second depth separable convolution layer is 32, the convolution kernel size is 3×3, and the step length is 2; feature map C9 is 32×32 for 32 channels; inputting the feature map C9 into a fourth standard convolution layer for convolution operation to obtain a feature map C10, wherein the number of convolution kernels of the fourth standard convolution layer is 32, the size of the convolution kernels is 1 multiplied by 1, and the step length is 1; feature map C10 is 32×32 for 32 channels; inputting the feature map C10 into a fourth normalized activation layer to perform batch normalization and activation operation to obtain a feature map C11; the feature map C11 is 32×32 of 32 channels.

III, inputting the feature map C8 and the feature map C11 into a tensor splicing layer for splicing to obtain a feature map C12, wherein the feature map C12 is 32 multiplied by 32 of 64 channels.

In this embodiment, the residual module includes a second standard convolution layer, a second normalized activation layer, a first depth separable convolution layer, a first batch of normalized layers, a third standard convolution layer, a third normalized activation layer, a second depth separable convolution layer, a second batch of normalized layers, a fourth standard convolution layer, a fourth normalized activation layer, and a tensor stitching layer.

Step S33: the feature map C12 is input to the attention module for feature extraction, and a feature map C25 is obtained.

As shown in fig. 3, step S33 specifically includes:

(1) Inputting a feature map C12 by using the attention input layer; the feature map C12 is 32×32 of 64 channels.

(2) Inputting the characteristic diagram C12 into a vertical average pooling layer for vertical pooling to obtain a characteristic diagram C13, wherein the characteristic diagram C13 is 32 multiplied by 1 of 64 channels; inputting the characteristic diagram C13 into a first dimension arrangement layer to perform dimension conversion of the horizontal and channel to obtain a characteristic diagram C16, wherein the characteristic diagram C16 is 64 multiplied by 1 of 32 channels; inputting the characteristic diagram C16 into a sixth standard convolution layer for convolution operation to obtain a characteristic diagram C19, wherein the number of convolution kernels of the sixth standard convolution layer is 32, the size of the convolution kernels is 3 multiplied by 1, the step length is 1, and the characteristic diagram C19 is 64 multiplied by 1 of 32 channels; inputting the feature map C19 into a first activation function layer for activation operation to obtain a feature map C21, wherein the activation function in the first activation function layer adopts a Sigmoid activation function, and the feature map C21 is 64 multiplied by 1 of 32 channels; the feature map C21 is input to a third three-dimensional arrangement layer to perform the dimensional transformation of the horizontal and channel, so as to obtain a feature map C23, wherein the feature map C23 is 32×1 of 64 channels.

(3) Inputting the characteristic diagram C12 into a horizontal average pooling layer for horizontal pooling to obtain a characteristic diagram C14, wherein the characteristic diagram C14 is 1 multiplied by 32 of 64 channels; inputting the feature map C14 into a second dimension arrangement layer to perform dimension conversion of the vertical and channel to obtain a feature map C17, wherein the feature map C17 is 1 multiplied by 64 of the 32 channels; inputting the characteristic diagram C17 into a seventh standard convolution layer for convolution operation to obtain a characteristic diagram C20, wherein the number of convolution kernels of the seventh standard convolution layer is 32, the size of the convolution kernels is 1 multiplied by 3, the step length is 1, and the characteristic diagram 20 is 1 multiplied by 64 of 32 channels; inputting the feature map C20 into a second activation function layer for activation operation to obtain a feature map C22, wherein the activation function in the second activation function layer adopts a Sigmoid activation function, and the feature map C22 is 1 multiplied by 64 of 32 channels; the feature map C22 is input to a fourth dimension arrangement layer to perform dimension conversion of vertical and channel, so as to obtain a feature map C24, wherein the feature map C24 is 1×32 of 64 channels.

(4) Inputting the feature map C12 into a fifth standard convolution layer for convolution operation to obtain a feature map C15, wherein the number of convolution kernels of the fifth standard convolution layer is 32, the size of the convolution kernels is 1 multiplied by 1, the step length is 1, and the feature map 15 is 32 multiplied by 32 of 64 channels; the feature map C15 is input to a fifth normalized activation layer for batch normalization and activation operation, so as to obtain a feature map C18, wherein the feature map C18 is 32×32 of 64 channels.

(5) The feature maps C23, C24 and C18 are input to the element multiplication layer to be subjected to element multiplication, so that a feature map C25 is obtained, and the feature map C25 is 32×32 of 64 channels.

In this embodiment, the attention module includes an attention input layer, a vertical average pooling layer, a horizontal average pooling layer, a fifth standard convolution layer, a first dimension arrangement layer, a second dimension arrangement layer, a fifth normalized activation layer, a sixth standard convolution layer, a seventh standard convolution layer, a first activation function layer, a second activation function layer, a third dimension arrangement layer, a fourth dimension arrangement layer, and an element multiplication layer.

Step S34: the feature map C25 is input to a second standard convolution module for feature extraction, and a feature map C28 is obtained.

Inputting the feature map C25 into an eighth standard convolution layer for convolution operation to obtain a feature map C26, wherein the number of convolution kernels of the second standard convolution layer is 128, the size of the convolution kernels is 3 multiplied by 3, and the step length is 2; feature map C26 is 16×16 for 128 channels; inputting the feature map C26 into a sixth normalized activation layer for batch normalization and activation operation to obtain a feature map C27; feature map C27 is 16×16 for 128 channels; inputting the characteristic diagram C27 into a second maximum pooling layer for maximum pooling operation to obtain a characteristic diagram C28; the size of the pooling window of the second maximum pooling layer is 3 multiplied by 3, and the step length is 2; feature map C28 is 8 x 8 for 128 channels.

In this embodiment, the second standard convolution module includes an eighth standard convolution layer, a sixth normalized activation layer, and a second maximum pooling layer.

Step S35: the feature map C28 is respectively subjected to numerical calculation to obtain a maximum value, a minimum value and an average value, and the one-dimensional image data includes the maximum value, the minimum value and the average value, specifically including:

1) The average value of the feature map C28 is calculated, the feature map C28 is 8×8 of 128 channels, and represents that the feature map is composed of 128 8×8 feature maps, for each feature map, the average value of all elements of the feature map can be calculated, and then the average value of all feature maps is averaged again, so as to obtain an average value representing the image feature.

2) The maximum value of the feature map C28 is calculated, the feature map C28 is 8×8 of 128 channels, and represents a composition of 128 8×8 feature maps, for each feature map, the maximum value of all elements thereof can be calculated, and then the maximum value is selected from the maximum values of all feature maps, so as to obtain a maximum value representing the image feature.

3) The minimum value of the feature map C28 is calculated, the feature map C28 is 8×8 of 128 channels, and represents that the feature map is composed of 128 8×8 feature maps, for each feature map, the minimum value of all elements of the feature map can be calculated, and then the minimum value is selected from the minimum values of all feature maps, so that a minimum value representing the features of the image is obtained.

In this embodiment, the satellite image network specifically includes a first standard convolution module, a residual module, an attention module, and a second standard convolution module.

Step S4: carrying out data processing operation on the one-dimensional image data and the time sequence data to obtain power load sequence data, wherein the method specifically comprises the following steps of:

step S41: and respectively carrying out data normalization processing on the maximum value, the minimum value and the average value in the one-dimensional image data to obtain a standard maximum value, a standard minimum value and a standard average value.

Step S42: and aligning the standard maximum value, the standard minimum value and the standard average value with the time sequence data according to the corresponding time stamps to obtain the power load sequence data.

Step S5: and respectively constructing a time sequence data set and a power load data set based on the time sequence data and the power load sequence data, wherein the time sequence data set is divided into a time sequence training set and a time sequence test set according to a certain proportion, and the power load data set is divided into a power load training set and a power load test set according to a certain proportion.

The method comprises the steps of respectively converting time sequence data and power load sequence data into supervision data, generating a data set based on a sliding window, selecting a certain number of load data variables at the moment to predict load data at the next moment, acquiring a large number of data in a sliding window mode to combine and generate the data set, wherein the width of the sliding window determines the time range of historical characteristic data for prediction. This width is chosen to take into account both prediction accuracy and computational efficiency.

If the width of the sliding window is too short, meaning that the input contains less historical characteristic data, some key information may be omitted, so that learning and training of a prediction model are affected; the predictive model may not capture long-term trends and periodic changes, resulting in inaccurate prediction results.

However, if the width of the sliding window is too long, meaning that the amount of input data for training the model increases, the training time and the computational complexity also increase accordingly; too long a sliding window may introduce too much history information, some of which may be irrelevant to the current predictions, resulting in the model training process becoming redundant and time consuming.

In this embodiment, data from 2020 to 2022 is selected. The acquisition period is 1 time/h, and 24 sampling points are arranged every day. The experimental training set is data in 2020 and 2021, and the test set is data in 2022. The invention selects the historical data of 72 sampling points (3 days) as input and the load data of 24 hours in the future as output.

Step S6: and inputting the time sequence training set and the electric load training set into a twin network to perform electric load prediction, so as to obtain an initial electric load prediction model.

As shown in fig. 4, step S6 specifically includes:

Step S61: inputting the time sequence training set into a first branch of a twin network to perform feature extraction to obtain time sequence features, wherein the method specifically comprises the following steps:

and inputting the time sequence training set into a first residual error module to perform expansion causal convolution operation to obtain a characteristic diagram T1.

And inputting the characteristic diagram T1 into a second residual error module to perform expansion causal convolution operation, so as to obtain a time sequence characteristic.

In this embodiment, the first branch includes a first residual module and a second residual module.

Step S62: inputting the electric load training set into a second branch of the twin network for feature extraction to obtain electric load features, wherein the method specifically comprises the following steps:

and inputting the time sequence training set into a first residual error module to perform expansion causal convolution operation to obtain a characteristic diagram T2.

And inputting the characteristic diagram T2 into a second residual error module to perform expansion causal convolution operation, so as to obtain the power load characteristic.

In this embodiment, the second branch includes a first residual module and a second residual module.

Step S63: and carrying out feature fusion on the time sequence features and the power load features to obtain power load fusion features.

Step S64: and sequentially inputting the power load fusion characteristics into the first full-connection layer and the second full-connection layer for prediction to obtain an initial power load prediction model.

Step S7: inputting the time sequence test set and the power load test set into an initial power load prediction model for testing, and outputting a final power load prediction model until the set conditions are met, wherein the method specifically comprises the following steps of:

step S71: and inputting the time sequence test set and the power load test set into an initial power load prediction model for testing to obtain a predicted power load.

Step S72: the predicted power load and the actual true value are calculated by adopting a mean square error (mean square error, MSE) loss function, and the formula is as follows:

in the method, in the process of the invention,indicate->Power load predictive value corresponding to time, +.>Indicate->And the real value of the power load corresponding to the moment.

Step S723: and judging whether the loss is smaller than a first threshold, outputting a final power load prediction model if the loss is smaller than the first threshold, and returning to the step S6 if the loss is larger than or equal to the first threshold.

Example 2

As shown in fig. 5, the present invention also provides a twin network-based power load prediction system, comprising:

the power data acquisition module 10 is used for acquiring historical power load data and historical power satellite images.

The feature engineering processing module 20 is configured to perform feature engineering processing on the historical power load data to obtain time sequence data.

The power image module 30 is configured to input a historical power satellite image into a satellite image network for feature extraction, obtain a feature map C28, and perform data conversion on the feature map to obtain one-dimensional image data.

The data processing module 40 is configured to perform a data processing operation on the one-dimensional image data and the time-series data to obtain power load sequence data.

The data set construction module 50 constructs a time series data set and a power load data set based on the time series data and the power load sequence data, respectively, the time series data set is divided into a time series training set and a time series test set according to a certain proportion, and the power load data set is divided into a power load training set and a power load test set according to a certain proportion.

The initial prediction module 60 is configured to input the time sequence training set and the power load training set into the twin network, and perform power load prediction to obtain an initial power load prediction model.

The test module 70 is configured to input the time sequence test set and the power load test set into the initial power load prediction model for testing, and output a final power load prediction model until a set condition is satisfied.

As an alternative embodiment, the power image module 30 of the present invention specifically includes:

The first standard convolution sub-module is used for inputting the historical electric satellite image into the first standard convolution module for feature extraction to obtain a feature map C3.

And the residual sub-module is used for inputting the characteristic diagram C3 into the residual module for characteristic extraction to obtain a characteristic diagram C12.

And the attention sub-module is used for inputting the feature map C12 into the attention module for feature extraction to obtain a feature map C25.

And the second standard convolution sub-module is used for inputting the characteristic diagram C25 into the second standard convolution module for characteristic extraction to obtain a characteristic diagram C28.

The numerical calculation sub-module is used for respectively carrying out numerical calculation on the feature map C28 to obtain a maximum value, a minimum value and an average value, and the one-dimensional image data comprises the maximum value, the minimum value and the average value.

As an alternative embodiment, the attention sub-module of the present invention specifically includes:

the attention input unit inputs the feature map C12 using the attention input layer.

The vertical average pooling unit is configured to input the feature map C12 to the vertical average pooling layer for vertical pooling to obtain a feature map C13, input the feature map C13 to the first dimension arrangement layer for dimension transformation to obtain a feature map C16, input the feature map C16 to the sixth standard convolution layer for convolution operation to obtain a feature map C19, input the feature map C19 to the first activation function layer for activation operation to obtain a feature map C21, and input the feature map C21 to the third dimension arrangement layer for dimension transformation to obtain a feature map C23.

The horizontal average pooling unit is configured to input the feature map C12 to the horizontal average pooling layer for horizontal pooling to obtain a feature map C14, input the feature map C14 to the second dimension arrangement layer for dimension transformation to obtain a feature map C17, input the feature map C17 to the seventh standard convolution layer for convolution operation to obtain a feature map C20, input the feature map C20 to the second activation function layer for activation operation to obtain a feature map C22, and input the feature map C22 to the fourth dimension arrangement layer for dimension transformation to obtain a feature map C24.

And the fifth convolution normalized activation unit inputs the feature map C12 to a fifth standard convolution layer to carry out convolution operation to obtain a feature map C15, and inputs the feature map C15 to a fifth normalized activation layer to carry out batch normalization and activation operation to obtain a feature map C18.

The element multiplication unit inputs the feature maps C23, C24, and C18 to the element multiplication layer to perform element multiplication, and obtains a feature map C25.

As an alternative embodiment, the initial prediction module of the present invention specifically includes:

and the first branch submodule is used for inputting the time sequence training set into a first branch of the twin network to perform feature extraction so as to obtain time sequence features.

And the second branch sub-module is used for inputting the electric load training set into a second branch of the twin network to perform feature extraction so as to obtain electric load features.

And the characteristic fusion sub-module is used for carrying out characteristic fusion on the time sequence characteristic and the power load characteristic to obtain the power load fusion characteristic.

And the characteristic prediction sub-module is used for predicting the fusion characteristic of the power load to obtain an initial power load prediction model.

The above description is only of the preferred embodiments of the present invention and is not intended to limit the present invention, but various modifications and variations can be made to the present invention by those skilled in the art. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A method of predicting power load based on a twin network, the method comprising:

2. The twin network-based power load prediction method according to claim 1, wherein the step of inputting the historical power satellite image into a satellite image network to perform feature extraction to obtain a feature map C28, and performing data conversion on the feature map to obtain one-dimensional image data specifically comprises:

3. The twin network-based power load prediction method according to claim 2, wherein the inputting the feature map C12 to the attention module performs feature extraction to obtain a feature map C25, specifically includes:

Inputting the feature map C12 using the attention input layer;

4. The twin network-based power load prediction method according to claim 2, wherein the performing a data processing operation on the one-dimensional image data and the time series data to obtain power load sequence data specifically comprises:

5. The twin network-based power load prediction method according to claim 1, wherein the inputting the time sequence training set and the power load training set into the twin network performs power load prediction to obtain an initial power load prediction model, and specifically comprises:

6. The method for predicting the power load based on the twin network as defined in claim 5, wherein the step of inputting the time sequence training set to the first branch of the twin network to perform feature extraction to obtain time sequence features specifically comprises:

7. A twin network-based electrical load prediction system, the system comprising:

8. The twin network-based power load prediction system of claim 7, wherein the power image module specifically comprises:

the residual sub-module is used for inputting the characteristic diagram C3 into the residual sub-module for characteristic extraction to obtain a characteristic diagram C12;

the attention submodule is used for inputting the feature map C12 into the attention submodule for feature extraction to obtain a feature map C25;

the second standard convolution sub-module is used for inputting the feature map C25 into the second standard convolution sub-module to perform feature extraction to obtain a feature map C28;

and the numerical calculation sub-module is used for respectively carrying out numerical calculation on the characteristic map C28 to obtain a maximum value, a minimum value and an average value.

9. The twin network based power load prediction system of claim 8, wherein the attention submodule comprises an attention input layer, a vertical average pooling layer, a horizontal average pooling layer, a fifth standard convolution layer, a first dimension arrangement layer, a second dimension arrangement layer, a fifth normalized activation layer, a sixth standard convolution layer, a seventh standard convolution layer, a first activation function layer, a second activation function layer, a third dimension arrangement layer, a fourth dimension arrangement layer, and an element multiplication layer, the attention submodule specifically comprising:

10. The twin network-based power load prediction system of claim 8, wherein the initial prediction module specifically comprises: