Short-term power load prediction method based on image domain feature extraction

Abstract

The invention discloses a short-term power load prediction method based on image domain feature extraction, and the method comprises the steps: obtaining and preprocessing multi-source time sequence data of a power load, and forming a one-dimensional power load historical information sequence data set; converting a one-dimensional time sequence in the data set from a time domain to an image domain by adopting a GAF algorithm to form a two-dimensional power load historical information image data set; inputting images in the two-dimensional power load historical information image data set into a ConvLSTM (Convolutional Long-Short-term Memory) model to extract daily and weekly period and time sequence characteristics of the power load and carry out short-term power load prediction; meanwhile, an extrusion-excitation (SE) channel attention mechanism is introduced into the ConvLSTM model so as to consider the nonlinear influence of the characteristic variable on the power load change. According to the method, the prior knowledge and the similar daily method are utilized to predict the logic of the power load, the power load prediction precision is improved, and automatic intelligent auxiliary reference is provided for power system dispatching decision making.

Description

Technical Field

[0001] The present invention belongs to the technical field of power load forecasting, and relates to a short-term power load forecasting method, and specifically to a short-term power load forecasting method based on image domain feature extraction. Background Art

[0002] Short-term power load forecasting mainly refers to the prediction of power load in the future (usually the next few hours to a week). Due to the special nature of electric energy, the power generation side and the power consumption side need to meet the real-time balance of supply and demand. Short-term power load forecasting helps to optimize the allocation of power resources and reduce unnecessary energy waste.

[0003] Through short-term power load forecasting, the power sector can formulate more reasonable power supply and dispatching plans, more accurately match power supply and demand, and reduce energy losses caused by over-power supply or under-power supply. Therefore, improving the accuracy of short-term power load forecasting is of great significance to the orderly operation of the power system. At present, some deep learning methods are widely used in power load feature extraction and short-term power load forecasting. Although deep learning methods can realize automatic feature extraction and end-to-end power load forecasting, the feature representation ability of one-dimensional time series in the time domain is limited. Deep learning methods often have problems such as overfitting, and the prediction effect is poor in practical application. Summary of the invention

[0004] The purpose of the present invention is to overcome the deficiencies of the prior art and to provide a short-term power load forecasting method based on image domain feature extraction, which can effectively realize accurate forecasting of different types of power loads.

[0005] To achieve the above object, the present invention adopts the following technical solution: A first aspect of an embodiment of the present invention provides a short-term power load forecasting method based on image domain feature extraction, the method specifically comprising the following steps: (1) Obtain and preprocess multi-source time series data of power load and its weather and date characteristic variables to form a one-dimensional power load historical information sequence data set. The weather characteristic variables and date characteristic variables include temperature, humidity, wind speed, working days, holidays and seasons; (2) using the Gram angular field algorithm to convert the one-dimensional time series in the one-dimensional power load historical information sequence data set obtained in step (1) from the time domain to the image domain, thereby forming a two-dimensional power load historical information image data set; (3) The images in the two-dimensional power load historical information image dataset obtained in step (2) are input into a convolutional long short-term memory model with a squeeze-excitation channel attention mechanism to extract the daily and weekly periodicity and time series characteristics of the power load and obtain the short-term power load forecasting results.

[0006] Furthermore, the step (1) specifically includes the following sub-steps: (1.1) Linear interpolation is used to fill missing values ​​in multi-source time series data; (1.2) Perform max-min normalization on the power load series and weather characteristic variables in the multi-source time series data after filling in missing values; (1.3) One-hot encoding of date feature variables in multi-source time series data after filling missing values; (1.4) The time granularity of the normalized weather characteristic variables and the one-hot encoded date characteristic variables are aligned with the sampling points of the power load series to form a one-dimensional power load historical information series dataset.

[0007] Specifically, the calculation formula of the linear interpolation method in step (1.1) is as follows: ; in, is the base time, As the base date, is a parameter, and its value is 3; Expressed as day The power load value sampled at every moment, Expressed as day The power load value sampled at every moment, Expressed as day The power load value sampled at each moment.

[0008] Specifically, the calculation formula for max-min normalization in step (1.2) is as follows: ; Among them, for the length Time Series , Indicates the sampling points, and ; and Represent the maximum and minimum values ​​in the time series respectively; is normalized to max-min The value of the interval.

[0009] Furthermore, the specific process of one-hot encoding in step (1.3) is as follows: the four seasons of spring, summer, autumn and winter are encoded as [0, 1, 2, 3] respectively; weekdays and weekends are encoded as [1, 0] respectively; holidays and non-holidays are encoded as [0, 1] respectively.

[0010] Furthermore, the step (2) specifically includes the following sub-steps: (2.1) The power load and its weather and date characteristic variables in the one-dimensional power load historical information sequence data set are converted to the interval [0,1]. The power load and its weather and date characteristic variables include temperature, humidity, wind speed, working days, holidays and seasons. Assume that the time series Length is , It is one of the sampling points, of which ,Will Mapping from the Cartesian coordinate system to the polar coordinate system, the calculation formula is as follows: ; in, represents the normalized power load sequence, It is The polar coordinate angle corresponding to each sampling point; (2.2) By considering the triangular sum between each sampling point in the polar coordinate system, the correlation of different time intervals of the time series can be encoded to form a Gram matrix, thereby imaging. The calculation formula is as follows: ; in, It is The polar coordinate angle corresponding to the sampling point is Indicates the number of time series sampling points contained in a picture; (2.3) According to the multi-periodicity of power load data, the images obtained by the Gram angular field algorithm are combined to obtain a two-dimensional power load historical information image dataset; since the power load has daily and weekly periodicity, the daily power load data within a week is regarded as a group, that is, a "video" consisting of 7 daily power load images, each image occupies 1 frame, and the specific representation is as follows: ; in, Indicates the The daily power load GAF ​​image is generated by taking the sampling points as the starting time. It represents a "video" consisting of 7 daily power load images; Similarly, and Indicates the The weather and date GAF images corresponding to the daily power load sampling points are generated for each sampling point at the start of measurement; and Respectively represent a "video" consisting of 7 weather and date images; and yes After integrating the three, we get a two-dimensional power load historical information image dataset, which is specifically expressed as follows: ; in, Represents the starting time of the two-dimensional power load historical information image data set Before moment and after Data at a moment; The time represents the starting time of the measurement Before The known historical information at each moment, power load ,date ,temperature ; Including the future Known information of the time to be predicted: Date ,temperature .

[0011] Furthermore, the squeeze-excitation channel attention mechanism in step (3) specifically includes the following sub-contents: (a) Global information aggregation: Perform global average pooling on the features of each input channel, compress the spatial dimension, and retain only the global information of each channel. The calculation formula is as follows: ; in, and is the length and width of the image, is the number of frames; the input is ,for A frame of is the number of feature channels; the output is ,in ; (b) Adaptive weight learning: The relationship between channels is captured through two fully connected layers and a nonlinear activation function. The calculation formula is as follows: ; in, is a dimension-raising matrix, For custom parameters, yes Activation function, yes Activation function, is a dimension reduction matrix; is the generated weight; (c) Feature recalibration: Re-adjust the input channel features with the new channel weights. The calculation formula is as follows: ;in, yes No. Feature map of each channel; It is the channel-weighted output of the original input.

[0012] Furthermore, the convolutional long short-term memory network in step (3) specifically includes the following contents: (a) Input gate: ; (b) Forget Gate: ; (c) Output gate: ; (d) Candidate memory element: ; (e) Update memory meta-state: ; (f) Update the hidden layer state: ; in, is a frame after the input feature is recalibrated; It is a convolution operation to extract features on the daily power load image; is the weight parameter, is the bias parameter, is the hidden layer state, represents the sigmoid activation function, represents the tanh activation function; It is the memory element of the weekly power load sequence, which can retain the time correlation of the power load within the week and the power load change trend information; The hidden state sequence of the frame After saving, the fully connected layer is input to obtain the power load forecast result.

[0013] Specifically, the power load prediction result in step (3) is expressed as: ; in, is the number of time points to be predicted, It is input by the convolutional long short-term memory network model. The model's trained parameters; Represents the SE-ConvLSTM model.

[0014] The present invention also discloses one or more memories and processors; the memory is coupled to the processor; wherein the memory is used to store program data, and the processor is used to execute the program data to implement the short-term power load forecasting method based on image domain feature extraction.

[0015] A second aspect of an embodiment of the present invention provides one or more memories and processors. The memory is coupled to the processor; wherein the memory is used to store program data, and the processor is used to execute the program data to implement the above-mentioned short-term power load forecasting method based on image domain feature extraction.

[0016] Compared with the prior art, the present invention has the following advantages and beneficial effects: 1. The present invention adopts the GAF algorithm to encode the one-dimensional time series into a two-dimensional power load image. The encoded image can reflect the characteristics of the power load from multiple angles, such as multi-periodicity and time sequence. Secondly, after encoding the one-dimensional sequence into a two-dimensional image, the successful paradigm in the field of computer vision can be used to better explore the characteristics of the power load.

[0017] 2. The present invention takes advantage of the SE attention mechanism in capturing global information of the channel, and uses the SE channel attention mechanism to adaptively assign weights to feature variables (such as temperature, humidity, wind speed, working days, holidays, seasons, etc.). Considering that the impact of feature variables on power load cannot be described only by linear relationships, after adding channel attention, nonlinear relationships can be used to describe the impact of feature variables on power load prediction results.

[0018] 3. The present invention uses the ConvLSTM model to fully extract the daily and weekly periodicity and time series characteristics of the power load. By inputting a video consisting of images of the power load history information of each day of the week into the ConvLSTM model, the convolution part of the model can extract the daily power load change trend characteristics from the daily power load data, while the long-term and short-term memory part of the model can retain the time correlation of the weekly power load data. By using the logic of predicting power load using prior knowledge and the similar day method, and incorporating known power load characteristic knowledge into the model, the accuracy of short-term power load prediction can be improved, with stronger generalization and robustness. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 is a flow chart of the method of the present invention; Figure 2 A flowchart of converting a time series into a GAF image according to the present invention; Figure 3 Four data input forms that can be obtained after using the GAF algorithm in the present invention; Figure 4 Schematic diagram of the SE channel attention module of the present invention; Figure 5 It is a unit diagram of FC-LSTM, MD-LSTM and ConvLSTM of the present invention; Figure 6 It is a curve chart of the short-term power load prediction result of "hourly system power load" based on image domain feature extraction of the present invention; Figure 7 It is a curve chart of the short-term power load prediction result of "15-minute commercial building power load" based on image domain feature extraction of the present invention; Figure 8 This is a curve chart of the short-term power load prediction results of "15-minute residential building power load" based on image domain feature extraction of the present invention. DETAILED DESCRIPTION

[0020] Exemplary embodiments will be described in detail herein, examples of which are shown in the accompanying drawings. When the following description refers to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present invention. Instead, they are merely examples of methods consistent with some aspects of the present invention as detailed in the appended claims.

[0021] The terms used in the present invention are only for the purpose of describing specific embodiments and are not intended to limit the present invention. The singular forms "a", "the" and "the" used in the present invention and the appended claims are also intended to include plural forms unless the context clearly indicates other meanings. It should also be understood that the term "and / or" used herein refers to and includes any or all possible combinations of one or more associated listed items.

[0022] The present invention will be further described in detail below in conjunction with the accompanying drawings and specific embodiments. In the absence of conflict, the features of the following embodiments and implementations can be combined with each other.

[0023] The power load is affected by many factors, and the main factors affecting the change of power load are weather factors and date factors. For weather factors, air conditioning power load is needed in hot summer, and lighting power load is needed when visibility is low due to rain; for date factors, the change of power load is affected by various time factors such as seasons, holidays, and commuting, and has complex periodicity such as days, weeks, and years. There are many types of power loads, which can be divided into rural power loads, urban residential power loads, industrial power loads, and commercial power loads, and each type of power load has different characteristics. Therefore, it is crucial to fully consider the multi-periodicity and timing of the power load itself as well as external influencing factors to achieve accurate short-term power load forecasting.

[0024] Based on this, the present invention proposes a short-term power load forecasting method based on image domain feature extraction, in which the SE channel attention module can adaptively assign weights to external influencing factors, fully considering their nonlinear influence on the power load value; the convolution part of ConvLSTM focuses on the daily periodicity of the power load sequence, extracts the changing trend characteristics of the daily power load, and the long and short-term memory part focuses on the weekly periodicity of the power load sequence, retaining the time correlation between the sampling points of the power load sequence.

[0025] like Figure 1 As shown, the present invention proposes a short-term power load forecasting method based on image domain feature extraction, which specifically includes the following steps: Step 1: Obtain multi-source time series data of power load and its characteristic variables such as weather and date (such as temperature, humidity, wind speed, working days, holidays, seasons, etc.) and preprocess them to form a one-dimensional power load historical information sequence data set.

[0026] Specifically, due to system failures or signal fluctuations, multi-source time series data inevitably have missing values. Linear interpolation can be used to fill in the missing values ​​of multi-source time series data. Linear interpolation is a simple interpolation method used to estimate the value of an unknown data point between two known data points. This method assumes that the change between two data points is linear, that is, the rate of change is constant. The calculation formula of linear interpolation is as follows: ; in, is the base time, As the base date, is a parameter, and its value is 3; Expressed as day The power load value sampled at every moment, Expressed as day The power load value sampled at every moment, Expressed as day The power load value sampled at each moment.

[0027] Since the power load series and weather characteristic variables have different dimensions and characteristic scales, they need to be normalized to make the numerical value range uniform. At the same time, normalization can also speed up the convergence of the model and improve the numerical stability of the model.

[0028] The power load series and weather characteristic variables in the multi-source time series data after filling the missing values ​​are normalized by max-min. The calculation formula is as follows: ; Among them, for the length Time Series , Indicates the sampling points, and ; and Represent the maximum and minimum values ​​in the time series respectively; is normalized to max-min The value of the interval.

[0029] Since the date feature variable is non-numerical data that cannot be directly input into the model, one-hot encoding can be used to represent each category with binary bits, and ensure that each category has a unique code. In this embodiment, the four seasons of spring, summer, autumn and winter are encoded as [0, 1, 2, 3] respectively; weekdays and weekends are encoded as [1, 0] respectively; holidays and non-holidays are encoded as [0, 1] respectively.

[0030] To ensure the consistency of the time intervals of weather features, date features, and power load data, for example, if the power load sequence is one data point per hour, the weather and date features also need to be aligned and integrated every hour. The time granularity of the normalized weather feature variables and the one-hot encoded date feature variables needs to be aligned with the sampling points of the power load sequence to form a one-dimensional power load historical information sequence data set.

[0031] Step 2: The Gram Angular Field (GAF) algorithm is used to convert the one-dimensional time series in the one-dimensional power load historical information sequence data set obtained in step 1 from the time domain to the image domain to form a two-dimensional power load historical information image data set.

[0032] GAF is a coding method that converts one-dimensional time series data into two-dimensional images, and is used to extract temporal dependencies and global information from time series data. GAF represents time series in the form of images through angle transformation and Gram matrix construction, so that the successful paradigm in the field of computer vision can be directly used for time series data analysis.

[0033] like Figure 2 As shown, in one embodiment, the specific implementation of step 2 includes the following sub-steps: Step 2.1: After converting the power load and its weather and date characteristic variables (including temperature, humidity, wind speed, working days, holidays, seasons, etc.) in the one-dimensional power load historical information sequence data set to the [0,1] interval, set the time series Length is , It is one of the sampling points, of which ,Will Mapping from the Cartesian coordinate system to the polar coordinate system, the calculation formula is as follows: ; in, represents the normalized power load sequence, It is The polar coordinate angle corresponding to each sampling point; (2.2) By considering the triangular sum between each sampling point in the polar coordinate system, the correlation of different time intervals of the time series can be encoded to form a Gram matrix, thereby imaging. The calculation formula is as follows: ; in, It is The polar coordinate angle corresponding to the sampling point is Indicates the number of time series sampling points contained in a picture; The GAF algorithm preserves the time dependency of the power load sequence ( Can indicate separation The relative correlation between the time steps of the data) and the temporal sequence (time increases from the upper left corner to the lower right corner of the image), and can be seen from the main diagonal At the same time, the data on the main diagonal of the image can be regarded as Within the range The data on the curve is mapped to Curve within the interval Therefore, the images generated by the GAF algorithm can provide a better data representation form on the model input side.

[0034] like Figure 3 As shown in the figure, after using the GAF algorithm to generate an image, four different data forms can be generated, represented by ①, ②, ③ and ④. ① represents a one-dimensional time series with a feature number of 1; ② represents a one-dimensional time series with a feature number of the number of matrix rows, where the columns of the matrix are used as features; ③ represents a two-dimensional matrix with a feature number of 1, where both the matrix rows and columns are considered as the time series propagation direction; ④ represents a two-dimensional image. The power load prediction accuracy of these four data input forms can be compared when performing short-term power load forecasting.

[0035] Step 2.3: According to the multi-periodicity of power load data, the images obtained by the GAF algorithm are combined to obtain a two-dimensional power load historical information image dataset. Since power load has daily and weekly periodicity, the daily power load data within a week is regarded as a group, which can be understood as a "video" composed of 7 daily power load images, each image occupies 1 frame, and the specific representation is as follows: ; in, Indicates the The daily power load GAF ​​image is generated by taking the sampling points as the starting time. It represents a "video" consisting of 7 daily power load images; Similarly, and Indicates the The weather and date GAF images corresponding to the daily power load sampling points are generated for each sampling point at the start of measurement; and Respectively represent a "video" consisting of 7 weather and date images; and yes After integrating the three, we get a two-dimensional power load historical information image dataset, which is specifically expressed as follows: ; in, Represents the starting time of the two-dimensional power load historical information image data set Before moment and after Data at a moment; The time represents the starting time of the measurement Before The known historical information at each moment, power load ,date ,temperature ; Including the future Known information of the time to be predicted: Date ,temperature .

[0036] Step 3: Input the images in the two-dimensional power load history information image dataset obtained in step 2 into the convolutional long short-term memory (ConvLSTM) model with squeeze-excitation (SE) channel attention mechanism to obtain the short-term power load forecasting results.

[0037] The SE module is a mechanism that enhances the network's expressive power by adaptively adjusting the importance of each channel. It can significantly enhance the generalization ability of the model because it dynamically adjusts feature expression by learning global channel dependencies. This mechanism is more robust to different data distributions. ConvLSTM is a deep learning model that combines convolutional neural networks (CNNs) and long short-term memory networks (LSTMs). It is particularly suitable for processing data with spatiotemporal dependencies. Compared with standard LSTMs or traditional CNNs, ConvLSTMs are more efficient in modeling long time series and more sensitive to local spatial features.

[0038] like Figure 4 As shown, the specific implementation process of step 3 is: First, perform global information aggregation. Perform global average pooling on the features of each input channel, compress the spatial dimension, and only retain the global information of each channel. The calculation formula is as follows: ; in, and is the length and width of the image, is the number of frames; the input is ,for A frame of is the number of feature channels; the output is ,in ; (b) Adaptive weight learning: The relationship between channels is captured through two fully connected layers and a nonlinear activation function. The calculation formula is as follows: ; in, is a dimension-raising matrix, For custom parameters, yes Activation function, yes Activation function, is a dimension reduction matrix; is the generated weight; (c) Feature recalibration: Re-adjust the input channel features with the new channel weights. The calculation formula is as follows: ;in, yes No. Feature map of each channel; It is the channel-weighted output of the original input.

[0039] The weighted data will be fed into the ConvLSTM network. FC-LSTM, MD-LSTM and ConvLSTM are for Figure 3 4 different data input forms of 3 networks are used to verify the superiority of ConvLSTM network, such as Figure 5 As shown in the figure (concat means concatenation). Data input forms ① and ② can be input into 1D FC-LSTM and 2D FC-LSTM respectively to extract time series features; MD-LSTM can simultaneously perform time series forward propagation and backtracking on the rows and columns of the two-dimensional matrix ③ to fully extract time series features. For data input form ④, ConvLSTM is used to fully extract spatiotemporal features to improve the accuracy of power load prediction.

[0040] The data information transmission process and update steps in the ConvLSTM unit are as follows: ① Input gate The input gate is one of the key mechanisms of ConvLSTM, which controls whether new information is added to the cell state. By dynamically adjusting the impact of input information, the input gate can help the network balance short-term input and long-term memory in complex timing tasks, making ConvLSTM more stable and flexible. The input gate can be expressed as: ; ② Forget Gate Responsible for dynamically controlling memory updates to prevent outdated information from interfering with the network. In ConvLSTM, through convolution operations, the forget gate not only retains the dependencies of the time series, but also captures local spatial features. The forget gate can be expressed as: ; ③Output gate The output gate is a key component in ConvLSTM, responsible for passing the relevant information of the cell state through the hidden state. The output gate can be expressed as: ; ④Candidate memory element Candidate memory cells can generate potential memory update content and provide the network with a rich expression of the current input. Candidate memory cells can be expressed as: ; (e) Update memory meta-state: ; (f) Update the hidden layer state: ; in, is a frame after the input feature is recalibrated; It is a convolution operation to extract features on the daily power load image; is the weight parameter, is the bias parameter, is the hidden layer state, represents the sigmoid activation function, represents the tanh activation function; It is the memory element of the weekly power load sequence, which can retain the time correlation of the power load within the week and the power load change trend information; The hidden state sequence of the frame After saving, the fully connected layer is input to obtain the power load forecast result.

[0041] Memory Metastate and hidden state The update of is completed through the collaboration of the input gate, forget gate, output gate and candidate memory cells. The process of memory cell update can be expressed as: ; The process of updating the hidden layer state can be expressed as: ; in, It is the memory element of the weekly power load sequence, which can retain the time correlation of the power load within the week and the power load change trend information. The hidden state sequence of the frame After saving, the data can be input into the fully connected layer to obtain the power load forecasting results, or used for other tasks.

[0042] The two-dimensional power load historical information image dataset is divided into training set, validation set and test set in a ratio of 8:1:1. During training, the model optimizer, loss function, learning rate and number of iterations are set to Adam, MSE, 0.001 and 200. After reaching the number of iterations, the model parameters are fixed, the test set is input into the model, and the short-term power load forecasting is completed. The process can be expressed as follows: ; in, is the number of time points to be predicted, is the model input, are the parameters of the prediction model.

[0043] To evaluate the accuracy of the prediction model, the power load prediction indicators include: MAE (Mean Absolute Error), RMSE (Root Mean Square Error), MAPE (Mean Absolute Percentage Error), and (R-Square, coefficient of determination), which is calculated as follows: ; ; ; ; in, It is The corresponding predicted value of samples, It is The true value corresponding to the sample is is the mean of the true values ​​of all samples to be predicted. The larger it is, the smaller the MAE, RMSE and MAPE are, and the better the prediction performance of the model is; otherwise, the worse the prediction performance of the model is. The values ​​of MAE and RMSE do not change with the numerical magnitude of the data set. On the contrary, the larger the numerical magnitude of the data set, the larger the values ​​of MAE and RMSE.

[0044] In order to verify the effectiveness and accuracy of the short-term power load forecasting method based on image domain feature extraction proposed in this invention, three types of power load data with different time granularity are used for simulation, where data set 1 comes from the system-level power load of the ISO New England control area, and the time interval of the power load data is 1h. The acquired data include hourly weather forecasts and power load values ​​from 2003 to 2014. Data sets 2 and 3 include 15-minute resolution power load values ​​of all major residential and commercial buildings in all climate zones in the United States in 2018, as well as dry-bulb temperature, humidity, wind speed and wind direction reported by weather stations in each state.

[0045] In order to better demonstrate the performance of the proposed method, the proposed SE-ConvLSTM model is compared with three LSTM-based short-term power load forecasting models: 1D SE-FC-LSTM, 2D SE-FC-LSTM and SE-MD-LSTM on the test sets of three power load data. The results of the comprehensive comparison of the performance of each model are shown in Table 1.

[0046] Table 1

[0047] The bold part in Table 1 represents the best performing model in a class of data sets. As shown in Table 1, the proposed method is superior to other methods in terms of MAE, RMSE, MAPE and indicators. The indices are all greater than 0.9. Compared with the 1D SE-FC-LSTM corresponding to the input without GAF algorithm encoding, the proposed method has a maximum reduction of 72.43% in the MAPE index, which effectively improves the accuracy of short-term power load forecasting.

[0048] Figure 6 , Figure 7 and Figure 8 The short-term power load forecasting result curves of "hourly system power load", "15-minute commercial building power load" and "15-minute residential building power load" based on image domain feature extraction proposed by the present invention are respectively shown. As can be seen from the figure, the short-term power load forecasting method based on image domain feature extraction proposed by the present invention can provide accurate short-term power load forecasting.

[0049] Accordingly, the present application also provides an electronic device, comprising: one or more memories and a processor. The memory is coupled to the processor; wherein the memory is used to store program data, and the processor is used to execute the program data to implement the above-mentioned short-term power load forecasting method based on image domain feature extraction.

[0050] Those skilled in the art will readily appreciate other embodiments of the present application after considering the description and practicing the contents disclosed herein. The present application is intended to cover any variations, uses or adaptations of the present application, which follow the general principles of the present application and include common knowledge or customary technical means in the art that are not disclosed in the present application. The description and examples are intended to be exemplary only.

[0051] It should be understood that the present application is not limited to the exact construction that has been described above and shown in the drawings, and that various modifications and changes may be made without departing from the scope thereof.

Claims

1. A short-term power load forecasting method based on image domain feature extraction, characterized in that: The method specifically comprises the following steps: (1) Obtain and preprocess multi-source time series data of power load and its weather and date characteristic variables to form a one-dimensional power load historical information sequence data set. The weather characteristic variables and date characteristic variables include temperature, humidity, wind speed, working days, holidays and seasons; (2) using the Gram angular field algorithm to convert the one-dimensional time series in the one-dimensional power load historical information sequence data set obtained in step (1) from the time domain to the image domain, thereby forming a two-dimensional power load historical information image data set; (3) The images in the two-dimensional power load historical information image dataset obtained in step (2) are input into a convolutional long short-term memory model with a squeeze-excitation channel attention mechanism to extract the daily and weekly periodicity and time series characteristics of the power load and obtain the short-term power load forecasting results.

2. The short-term power load forecasting method based on image domain feature extraction according to claim 1 is characterized in that: The step (1) specifically includes the following sub-steps: (1.1) Linear interpolation is used to fill missing values ​​in multi-source time series data; (1.2) Perform max-min normalization on the power load series and weather characteristic variables in the multi-source time series data after filling in missing values; (1.3) One-hot encoding of date feature variables in multi-source time series data after filling missing values; (1.4) The time granularity of the normalized weather characteristic variables and the one-hot encoded date characteristic variables are aligned with the sampling points of the power load series to form a one-dimensional power load historical information series dataset.

3. The short-term power load forecasting method based on image domain feature extraction according to claim 2 is characterized in that: The calculation formula of the linear interpolation method in step (1.1) is as follows: ； in, is the base time, As the base date, is a parameter, and its value is 3; Expressed as day The power load value sampled at every moment, Expressed as day The power load value sampled at every moment, Expressed as day The power load value sampled at each moment.

4. The short-term power load forecasting method based on image domain feature extraction according to claim 2 is characterized in that: The calculation formula for max-min normalization in step (1.2) is as follows: ; Among them, for the length Time Series , Indicates the sampling points, and ; and Represent the maximum and minimum values ​​in the time series respectively; is normalized to max-min The value of the interval.

5. The short-term power load forecasting method based on image domain feature extraction according to claim 2 is characterized in that: The specific process of one-hot encoding in step (1.3) is as follows: the four seasons of spring, summer, autumn and winter are encoded as [0, 1, 2, 3] respectively; weekdays and weekends are encoded as [1, 0] respectively; holidays and non-holidays are encoded as [0, 1] respectively.

6. The short-term power load forecasting method based on image domain feature extraction according to claim 1 is characterized in that: The step (2) specifically includes the following sub-steps: (2.1) The power load and its weather and date characteristic variables in the one-dimensional power load historical information sequence data set are converted to the interval [0,1]. The power load and its weather and date characteristic variables include temperature, humidity, wind speed, working days, holidays and seasons. Assume that the time series Length is , It is one of the sampling points, of which ,Will Mapping from the Cartesian coordinate system to the polar coordinate system, the calculation formula is as follows: ； in, represents the normalized power load sequence, It is The polar coordinate angle corresponding to the sampling points; (2.2) By considering the triangular sum between each sampling point in the polar coordinate system, the correlation of different time intervals of the time series can be encoded to form a Gram matrix, thereby imaging. The calculation formula is as follows: ； in, It is The polar coordinate angle corresponding to the sampling point is Indicates the number of time series sampling points contained in a picture; (2.3) According to the multi-periodicity of power load data, the images obtained by the Gram angular field algorithm are combined to obtain a two-dimensional power load historical information image dataset; since the power load has daily and weekly periodicity, the daily power load data within a week is regarded as a group, that is, a "video" consisting of 7 daily power load images, each image occupies 1 frame, and the specific representation is as follows: ； in, Indicates the The daily power load GAF ​​image is generated by taking the sampling points as the starting time. Represents a "video" consisting of 7 daily power load images; Similarly, and Indicates the The weather and date GAF images corresponding to the daily power load sampling points are generated for each sampling point at the start of measurement; and Respectively represent a "video" consisting of 7 weather and date images; and yes After integrating the three, we get a two-dimensional power load historical information image dataset, which is specifically expressed as follows: ； in, Represents the starting time of the two-dimensional power load historical information image data set Before moment and after Data at a certain moment; The time represents the starting time of the measurement Before The known historical information at each moment, power load ,date ,temperature ; Including the future Known information of the time to be predicted: Date ,temperature .

7. The short-term power load forecasting method based on image domain feature extraction according to claim 1 is characterized in that: The squeeze-excitation channel attention mechanism in step (3) specifically includes the following sub-contents: (a) Global information aggregation: Perform global average pooling on the features of each input channel, compress the spatial dimension, and retain only the global information of each channel. The calculation formula is as follows: ； in, and is the length and width of the image, is the number of frames; the input is ,for A frame of is the number of feature channels; the output is ,in ; (b) Adaptive weight learning: The relationship between channels is captured through two fully connected layers and a nonlinear activation function. The calculation formula is as follows: ； in, is a dimension-raising matrix, For custom parameters, yes Activation function, yes Activation function, is a dimension reduction matrix; is the generated weight; (c) Feature recalibration: Re-adjust the input channel features with the new channel weights. The calculation formula is as follows: ;in, yes No. Feature map of each channel; It is the channel-weighted output of the original input.

8. The short-term power load forecasting method based on image domain feature extraction according to claim 1 is characterized in that: The convolutional long short-term memory network in step (3) specifically includes the following contents: (a) Input gate: ； (b) Forget Gate: ； (c) Output gate: ； (d) Candidate memory element: ； (e) Update memory meta-state: ； (f) Update the hidden layer state: ； in, is a frame after the input feature is recalibrated; It is a convolution operation to extract features on the daily power load image; is the weight parameter, is the bias parameter, is the hidden layer state, represents the sigmoid activation function, represents the tanh activation function; It is the memory element of the weekly power load sequence, which can retain the time correlation of the power load within the week and the power load change trend information; The hidden state sequence of the frame After saving, the fully connected layer is input to obtain the power load forecast result.

9. The short-term power load forecasting method based on image domain feature extraction according to claim 1 is characterized in that: The power load prediction result in step (3) is expressed as: ； in, is the number of time points to be predicted, It is input by the convolutional long short-term memory network model. The model's trained parameters; Represents the SE-ConvLSTM model.

10. The short-term power load forecasting method based on image domain feature extraction according to claim 1 is characterized in that: include: One or more memories and processors; the memory is coupled to the processor; wherein the memory is used to store program data, and the processor is used to execute the program data to implement the short-term power load forecasting method based on image domain feature extraction as described in any one of claims 1-9 above.

Images