Power load prediction method based on multi-modal feature extraction and decomposition fusion

Abstract

The invention discloses a power load prediction method based on multi-modal feature extraction and decomposition fusion, and relates to the technical field of data analysis. According to the method, a complex load time sequence is decomposed into a high-frequency component and a low-frequency component by using a decomposition algorithm, and the technology of carrying out feature extraction and prediction by combining a deep learning model is widely concerned. According to the method, multi-scale information in the data can be better captured, and the prediction precision and adaptability of the model are improved.

Description

Technical Field

[0001] The invention discloses a method for predicting electric load based on multi-modal feature extraction and decomposition fusion, and relates to the technical field of data analysis. Background Art

[0002] With the development of power systems and the popularization of smart grids, power load forecasting plays a vital role in optimizing power dispatching, improving grid operation efficiency, and reducing energy waste. In order to improve the accuracy of load forecasting, data-driven methods have gradually become a research hotspot.

[0003] Traditional load forecasting methods mostly rely on a single statistical model or machine learning model, such as ARIMA, SVR, etc. Although these methods are effective in some scenarios, they have limited performance when processing load data with complex nonlinear and time-varying characteristics. In addition, some existing deep learning models, such as GRU and LSTM, although they have advantages in capturing time series dependencies, lack independent modeling of high-frequency and low-frequency characteristics of load data, resulting in low accuracy when predicting complex load fluctuations. Existing technologies also often ignore the feature extraction of external factors such as meteorological data that are highly correlated with load, limiting the comprehensiveness and accuracy of the model. Summary of the invention

[0004] In view of the problems of the prior art, the present invention provides a method and device for power load forecasting based on multimodal feature extraction and decomposition fusion, and the technical solution adopted is: In a first aspect, a method for power load forecasting based on multimodal feature extraction and decomposition fusion is provided, the method comprising: S1, obtain time series data based on historical charge coincidence and meteorological data, and perform data cleaning; S2, according to the time series data, decomposing into a plurality of intrinsic mode functions and a residual term through the time series, and differentiating into a high-frequency component and a low-frequency component according to the frequency of the intrinsic mode function; S3, performing weight partitioning by using the intrinsic mode function according to the meteorological data; S4, processing the high frequency components by combining feature extraction with a parallel model; S5, capturing the nonlinear trend through a gradient boosting regression tree according to the low-frequency component; S6, adding the prediction results of the intrinsic mode functions and recombining them as a prediction value.

[0005] In some implementations, the S3 specifically includes: S31, based on the correlation between the meteorological data and the power consumption, evaluating the correlation of each meteorological data to the power load by using the Pearson correlation coefficient.

[0006] In some implementations, the S31 specifically includes: S311, assigning an initial weight to each meteorological factor of the meteorological data according to temperature, humidity and wind speed; S312, according to the Pearson correlation coefficient, the correlation between each meteorological data and the power consumption is analyzed by formula (2); the formula (2) is as follows: (2) In the formula express The average value of express The average value of S313, according to the load data, analyzing the correlation coefficient through each of the meteorological data, wherein the correlation coefficient includes the temperature correlation coefficient ; Humidity correlation coefficient ; Wind speed correlation coefficient ; According to the correlation coefficient, the expert initial weight is adjusted by formula (16); the formula (16) is as follows: (16) in, is the final desired weight.

[0007] In some implementations, the S3 further includes: S32, automatically adjusting through a neural network according to the correlation and the expert initial weight.

[0008] In some implementations, the S32 specifically includes: S321, performing model training through the neural network according to the expert initial weights and input data; S322, according to the model training, performing hidden layer processing and feature extraction according to the neural network, and adjusting the expert initial weights through a back propagation algorithm; S323, output a prediction value according to the output layer of the neural network , and the actual power consumption For comparison, the mean square error is used for measurement, and the formula is as follows: (18) in, is the load value predicted by the model; is the actual load value; N is the number of samples; S324, according to the back propagation algorithm, propagate the error forward layer by layer through the chain rule, and adjust the weights and biases of each layer of the neural network.

[0009] In some implementations, the S4 specifically includes: S41, capturing spatiotemporal characteristics through feature extraction according to the high-frequency component, and taking the decomposed high-frequency component and meteorological data as input, performing feature extraction to obtain channel features, spatial features and temporal features; S42, forming a feature vector by feature fusion according to the channel feature, spatial feature and temporal feature.

[0010] In a second aspect, an embodiment of the present invention provides a power load forecasting device based on multimodal feature extraction and decomposition fusion, the device comprising: The preprocessing module is used to obtain time series data based on historical charge coincidence and meteorological data, and perform data cleaning; A sequence decomposition module, used for decomposing the time series data into a plurality of intrinsic mode functions and a residual term through the time series, and differentiating into a high-frequency component and a low-frequency component according to the frequency of the intrinsic mode function; A weight distribution module, used for performing weight partitioning through the intrinsic mode function according to the meteorological data; A combined processing module, used for combining feature extraction with a parallel model according to the high frequency component; A low-frequency prediction module, used to capture nonlinear trends through a gradient boosting regression tree according to the low-frequency components; The prediction reconstruction module is used to add the prediction results of the intrinsic mode functions and recombine them as the prediction value.

[0011] In some implementations, the weight allocation module specifically includes: A construction unit is used to evaluate the relevance of each meteorological data to the electric load by using a Pearson correlation coefficient according to the relevance between the meteorological data and the electric power consumption; An adjustment unit is used to automatically adjust through a neural network according to the correlation and the expert initial weight.

[0012] In a third aspect, an embodiment of the present invention provides an electronic device, comprising a memory and a processor, wherein the memory is used to store one or more computer instructions, wherein when the one or more computer instructions are executed by the processor, the method described in the first aspect above is implemented.

[0013] In a fourth aspect, an embodiment of the present invention provides a computer storage medium, wherein the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, it is used to implement the method described in the first aspect.

[0014] One or more embodiments of the present invention can bring at least the following beneficial effects: The beneficial effects of the present invention are as follows: the method of the present invention uses a decomposition algorithm (such as EEMD) to decompose complex load time series into high-frequency and low-frequency components, and then combines the deep learning model to extract features and predict the technology, which has attracted widespread attention. This method can better capture multi-scale information in the data and improve the prediction accuracy and adaptability of the model.

[0015] The method of the present invention proposes a hybrid model combining EEMD decomposition, CNN and GBRT, and can better solve the above-mentioned shortcomings and improve the accuracy and generalization ability of load forecasting by extracting features and parallel predicting high-frequency and low-frequency components respectively. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0017] Figure 1 It is an experimental flow chart of a power load forecasting method based on multi-modal feature extraction and decomposition fusion provided by an embodiment of the present invention; Figure 2 It is a CBAM-CNN-LSTM flow chart provided by an embodiment of the present invention; Figure 3 is an LSTM flow chart provided by an embodiment of the present invention; Figure 4 It is a CBAM flow chart provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0018] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. The components of the embodiments of the present invention generally described and shown in the drawings here can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the drawings is not intended to limit the scope of the claimed invention, but merely represents the selected embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without making creative work belong to the scope of protection of the present invention.

[0019] First, some professional terms involved in this embodiment are explained: 1. Data normalization This experiment uses maximum and minimum normalization to preprocess the data. Maximum and minimum normalization uses the maximum and minimum values ​​in the data for standardization, so that the standardized data is between [0,1], thereby eliminating the adverse effects caused by abnormal value data. The formula is as follows: (1) in, is the normalized value, is the original data before normalization, is the maximum value of the sample data, is the minimum value of the sample data.

[0020] 2. Specific principle of Pearson correlation coefficient: The Pearson correlation coefficient is used to indicate the correlation between the measured value and the true value, and the correlation coefficient varies between -1 and 1. The closer it is to 1 or -1, the better the correlation is. The higher the correlation is, the more accurate the measured data is. Similarly, the closer the obtained data is to 0, the lower the correlation of the obtained values ​​is, and the obtained value is inaccurate. A coefficient value of 1 means that X and Y can be used to better represent the data as a straight line equation, all data can fall on the linear equation, and the linear equation is positively correlated. Similarly, a coefficient value of −1 also means that all data points fall on the straight line, but at this time X and Y are negatively correlated, and Y decreases as X increases. A value of 0 indicates that there is no correlation between them, and the data is not on the straight line. The calculation formula is shown in Formula 2: (2) In the formula express The average value of express The average value of .

[0021] 3. Specific principles of CBAM: CBAM is a simple and efficient feedforward convolutional neural network module. The module can infer the attention map from two different dimensions (channel and space) and multiply the attention map with the input feature map to achieve feature optimization of attention. Because CBAM is a lightweight universal module, it can be seamlessly integrated into any CNN architecture and can be trained end-to-end through basic CNN. Figure 1 This is the CBAM network structure diagram.

[0022] The output of the convolutional layer first passes through the channel attention module, and after obtaining the weight, it passes through the spatial attention module and finally performs the weighted calculation. The channel attention mechanism compresses the feature map in the spatial dimension to obtain a one-dimensional vector, and then it can be processed. When compressing in the spatial dimension, not only average pooling (Average Pooling) but also maximum pooling (Max Pooling) is considered. Average pooling and maximum pooling can be used to aggregate the spatial information of the feature map, send it to a shared network, compress the spatial dimension of the input feature map, and sum and merge element by element to generate a channel attention map. Average pooling has feedback for each pixel on the feature map, while when performing gradient back propagation calculation for maximum pooling, only the place with the largest response in the feature map has gradient feedback. The expression of the channel attention mechanism is shown in (3): (3) The spatial attention mechanism compresses the channel, and uses average pooling and maximum pooling in the channel dimension. The MaxPool operation extracts the maximum value from the channel, and the frequency of each extraction is the height multiplied by the width; the AvgPool operation extracts the channel on average, and the number of extractions is also the height multiplied by the width; then the feature maps extracted previously (the number of channels is 1) are merged to obtain a 2-channel feature map, as shown in formula (4): (4) Among them, σ is the sigmoid operation, 7*7 represents the size of the convolution kernel, and the 7*7 convolution kernel has a better effect than the 3*3 convolution kernel.

[0023] 4. CNN specific principles: Convolutional neural network (CNN) contains a feature extractor consisting of convolutional layer and subsampling layer (pooling layer). In the convolutional layer of convolutional neural network, a neuron is only connected to some neurons in the adjacent layers.

[0024] Convolutional layer: Each convolutional layer in a convolutional neural network consists of several convolutional units, and the parameters of each convolutional unit are optimized through the back-propagation algorithm. The purpose of the convolution operation is to extract different features of the input. The first convolutional layer may only extract some low-level features such as edges, lines, and corners. More layers of the network can iteratively extract more complex features from low-level features.

[0025] Pooling layer: Usually after the convolution layer, we get features with large dimensions. We cut the features into several regions, take their maximum or average value, and get new features with smaller dimensions.

[0026] Fully connected layer: combines all local features into global features to calculate the final score of each category.

[0027] 5. LSTM principle: Long short-term memory network (LSTM) is a recurrent neural network designed to overcome the gradient vanishing and exploding problems in long sequence training. The LSTM structure mainly includes: input gate, forget gate, output gate, and internal memory unit.

[0028] Among them, the forget gate is used to control whether to forget. In LSTM, it controls whether to forget the hidden cell state of the previous layer with a certain probability. The input content includes the hidden state h(t-1) of the previous sequence and the current sequence data xt. Through the sigmoid activation function, the output f(t) of the forget gate is obtained. At the same time, this output also represents the probability of forgetting the hidden cell state of the previous layer. The specific mathematical expression is shown in formula (5).

[0029] (5) Where Wf, Uf, bf are the coefficients and offsets of the linear relationship, and σ is the sigmoid activation function.

[0030] The input gate mainly processes the input of the current sequence position. It consists of two parts, including the sigmoid activation function and the tanh activation function. The outputs of the two are multiplied to update the cell state. The mathematical expressions are shown in equations (6)-(7).

[0031] (6) (7) Where Wi, Ui, bi, Wa, Ua, and ba are the coefficients and offsets of the linear relationship, and σ and tanh are activation functions.

[0032] The internal memory unit consists of two parts. The first part is the product of the previous cell state and the output of the forget gate, and the other part is the product of the input gate. The specific mathematical expression is shown in formula (8).

[0033] (8) Where ⊙ is the Hadamard product.

[0034] The output gate is mainly composed of two parts. The first part includes the hidden state of the previous sequence, the current sequence data and the sigmoid activation function; the second part consists of the hidden state and the tanh activation function. The mathematical expressions are shown in equations (9)-(10).

[0035] (9) (10) 6. The specific principle of EEMD algorithm: The essence of the EEMD method is a multiple empirical mode decomposition of superimposed Gaussian white noise. It uses the statistical characteristics of Gaussian white noise with uniform frequency distribution, and changes the extreme point characteristics of the signal by adding different white noises of the same amplitude each time. Then, the corresponding IMFs obtained by multiple EEMDs are averaged to offset the added white noise, thereby effectively suppressing the generation of modal aliasing. For example, the original input signal IMF1 is decomposed into IMF2-IMF7 components from high frequency to low frequency and the residual signal IMF8 ( ).

[0036] 7. Gradient Boosting Regression Tree (GBRT) Principle GBRT is an ensemble learning method that improves prediction accuracy by integrating multiple decision tree models into a powerful model. The core idea is to correct the errors of the previous model by gradually building new decision trees, and finally get a more accurate model. The following is a detailed introduction to the working principle of GBRT.

[0037] First, GBRT initializes a simple constant model , which is usually the mean of the target values ​​in the training data: (11) in, is the true target value, c is a constant, L is the loss function, and n is the number of samples of training data.

[0038] Next, GBRT continuously updates the model through a step-by-step iterative process. In each iteration, the goal of GBRT is to minimize the residual error of the loss function (i.e., the error between the previous round's predicted value and the true value) by fitting a new regression tree. The key steps are as follows: (1) Calculate the residual (gradient): In the mth iteration, calculate the residual between the predicted value of the current model and the true target value. The residual reflects the error size of the model's current data fit. The calculation of the residual is based on the gradient of the loss function: (12)

[0039] (2) Fitting a regression tree: using residuals As the target value, fit a new regression tree to predict these residuals. The purpose of the new regression tree is to find the best fit relationship between the input feature x and the residuals, thereby correcting the prediction errors of the previous model.

[0040] (3) Update the model: Add the new regression tree to the existing model and use the learning rate η to control the contribution of the new tree to the overall model. The model update formula is: (13) in, is the model after the mth iteration, is the fitted regression tree, is the learning rate (the value range is generally 0<< ≤1), which determines the contribution of each tree.

[0041] Repeat the iterations. By repeating the above process, the GBRT model will gradually improve until the preset number of iterations is reached or other stopping conditions are met.

[0042] (4) Loss function selection One advantage of GBRT is that it can choose different loss functions to adapt to different task requirements. Commonly used loss functions are: Mean Squared Error (MSE): Applicable to regression tasks, measuring the difference between the predicted value and the true value.

[0043] (14) Log-likelihood loss: Suitable for classification tasks, usually used for binary classification problems.

[0044] (15) Embodiment 1: Figure 1 An experimental flow chart of a power load forecasting method based on multimodal feature extraction and decomposition fusion is shown, Figure 1 As shown, a method for power load forecasting based on multimodal feature extraction and decomposition fusion, the method comprising: S1, obtain time series data based on historical charge coincidence and meteorological data, and perform data cleaning; S2, according to the time series data, decomposing into a plurality of intrinsic mode functions and a residual term through the time series, and differentiating into a high-frequency component and a low-frequency component according to the frequency of the intrinsic mode function; S3, performing weight partitioning by using the intrinsic mode function according to the meteorological data; S4, processing the high frequency components by combining feature extraction with a parallel model; S5, capturing the nonlinear trend through a gradient boosting regression tree according to the low-frequency component; S6, adding the prediction results of the intrinsic mode functions and recombining them as a prediction value.

[0045] First, according to S1, time series data are obtained from sources such as historical power load and meteorological data. Data cleaning is performed to handle missing values ​​and outliers, and data normalization (Min-Max normalization) is performed to ensure that the data input format is consistent.

[0046] Next, according to S2, the original time series is decomposed into multiple intrinsic mode functions (IMFs) using the time series decomposition EEMD method, each of which represents a component of different frequencies. Applying EEMD to the time series decomposes it into several IMF components and a residual term. According to the frequency of each IMF component, the components are divided into high-frequency components and low-frequency components. Generally speaking, lower-frequency IMFs reflect long-term trends, while high-frequency IMFs represent short-term fluctuations.

[0047] Next, follow S3, first, through S31, Specifically, S3 specifically includes: in order to reduce redundancy and reasonably allocate weights of different meteorological factors, a weighting method based on expert experience and data-driven is proposed.

[0048] Correlation analysis of weather data: Calculate the correlation between each weather data (such as temperature, humidity, wind speed, precipitation, etc.) and electricity consumption, and use the Pearson correlation coefficient to evaluate the correlation of each weather data to the electricity load.

[0049] Subjective weighting method: Based on expert experience, higher initial weights are given to weather data with higher relevance. This method is based on the subjective assessment of the importance of meteorological data by multiple experts based on their understanding of local meteorological conditions and load changes.

[0050] Specifically, the S31 specifically includes: S311, assigning an initial weight to each meteorological factor of the meteorological data according to temperature, humidity and wind speed; S312, according to the Pearson correlation coefficient, the correlation between each meteorological data and the power consumption is analyzed by formula (2); the formula (2) is as follows: (2) In the formula express The average value of express The average value of S313, according to the load data, analyzing the correlation coefficient through each of the meteorological data, wherein the correlation coefficient includes the temperature correlation coefficient ; Humidity correlation coefficient ; Wind speed correlation coefficient ; According to the correlation coefficient, the expert initial weight is adjusted by formula (16); the formula (16) is as follows: (16) in, is the final desired weight.

[0051] Then, through S32, automatic adjustment is performed through a neural network according to the correlation and the expert initial weight.

[0052] Specifically, the S32 specifically includes: According to S321, model training is performed through the neural network according to the expert initial weights and input data; First, the initial weights obtained based on correlation analysis and expert weighting method will be used as trainable parameters in the network. These weights will participate in model training together with the input data. The corresponding weight The initial weight value is calculated by the previous step. The input of the model is multiple weather data And the corresponding initial weight After the input layer of the neural network receives this data, it first weights the impact of the weather data: (17) in, represents weather data with weights, is the current weight, is the actual value of the weather data.

[0053] Next, according to S322, according to the model training, hidden layer processing and feature extraction are performed according to the neural network, and the expert initial weights are adjusted by the back propagation algorithm; After the input layer, the neural network usually contains one or more hidden layers to further extract complex features. Use network structures such as convolutional neural network (CNN) and long short-term memory network (LSTM) to process input data. The network structure includes the following parts: Convolutional layer (CNN): extracts spatial features from meteorological factors and load data, and is particularly suitable for capturing local information.

[0054] LSTM: used for time series data to capture the temporal dependencies of historical loads and meteorological factors.

[0055] Fully connected layer: Integrates the features extracted by the upper layer for the final prediction output.

[0056] During the training of the hidden layer, the network adjusts the weights through the backpropagation algorithm based on the output error of each layer.

[0057] Next, according to S323, a prediction value is output according to the output layer of the neural network. , and the actual power consumption For comparison, the mean square error is used for measurement, and the formula is as follows: (18) in, is the load value predicted by the model; is the actual load value; N is the number of samples; S324, during the back propagation process, the network will adjust the weights based on the prediction error The back propagation algorithm propagates the error forward layer by layer through the chain rule, adjusting the weights and biases of each layer to ensure that the network can gradually reduce the prediction error. The next step, according to S4, specifically includes: S41, capturing spatiotemporal characteristics through feature extraction according to the high-frequency component, and taking the decomposed high-frequency component and meteorological data as input, performing feature extraction to obtain channel features, spatial features and temporal features; Feature extraction is performed on the decomposed high-frequency components to capture their complex spatiotemporal characteristics. The decomposed high-frequency components and weather data are used as input to perform the following feature extractions: First, the decomposed high-frequency load data is input into the channel feature extraction CBAM module. The CBAM module extracts important channel features for the load data through the channel attention mechanism, suppresses irrelevant channels, and thus enhances the model's sensitivity to frequency changes.

[0058] The high-frequency load data and high-weight weather data after feature extraction are input into the spatial feature extraction CNN model to extract their spatial features. CNN captures the local correlation between data through convolution operations, especially the local change pattern of high-frequency data.

[0059] The high-frequency load data after feature extraction, the weather data with high weight, and the remaining weather data are input into the time feature extraction LSTM model. The LSTM model can capture short-term dependencies in time series and is suitable for processing time dynamic changes in high-frequency data.

[0060] S42, forming a feature vector by feature fusion according to the channel feature, spatial feature and temporal feature.

[0061] The channel features, spatial features, and temporal features extracted by CBAM, CNN, and LSTM are combined to form a complete feature vector. This is accomplished through full connection.

[0062] Next, according to S5, the low-frequency components have long-term trends, and GBRT (gradient boosted regression tree) is used to capture nonlinear trends.

[0063] Data input: Input the low-frequency IMF components into GBRT.

[0064] Model structure: GBRT is essentially an integrated model of multiple decision trees. New trees are constructed through the residuals of each round of iteration to improve the overall prediction accuracy.

[0065] Training: GBRT is trained, and parameters such as tree depth and learning rate can be adjusted through grid search or cross-validation.

[0066] Finally, according to S6, the forecast results of each IMF component are added together and recombined into the complete forecast value.

[0067] It can be seen from the above embodiments that, compared with the prior art, the advantages of the present application are: 1. Combination of EEMD decomposition with high- and low-frequency prediction: This method combines EEMD decomposition with deep learning models and integrated models to perform differentiated processing on high-frequency components and low-frequency components. High-frequency components are feature extracted using CBAM, CNN, and LSTM, focusing on capturing complex spatiotemporal dynamic changes; low-frequency components are captured through the GBRT model to capture long-term trends. This method can make full use of the characteristics of different frequency components to improve the accuracy of prediction and the generalization ability of the model.

[0068] 2. Combination of weight allocation and automatic adjustment of neural network: A weighting method based on expert experience and data-driven is proposed, combined with the automatic weight adjustment mechanism of neural network. By analyzing the correlation between weather data and load data, the weights of various weather factors are preliminarily allocated, and the weights are further automatically adjusted during the neural network training process to achieve accurate contribution analysis of weather data to load forecasting. This two-level weight optimization method effectively reduces redundancy and improves the stability and adaptability of the model.

[0069] 3. Parallel feature extraction and prediction fusion: A parallel feature extraction mechanism is proposed for high-frequency components. Different weather data are input into different models according to the weight of their correlation to extract data features. The channel, space and time features of CBAM, CNN and LSTM are combined to fully capture the complex characteristics of load data. Compared with the traditional single feature extraction method, it can more comprehensively characterize the fluctuation characteristics of high-frequency components, thereby enhancing the model's sensitivity to frequency changes.

[0070] Embodiment 2: In a second aspect, an embodiment of the present invention provides a power load forecasting device based on multimodal feature extraction and decomposition fusion, the device comprising: The preprocessing module is used to obtain time series data based on historical charge coincidence and meteorological data, and perform data cleaning; A sequence decomposition module, used for decomposing the time series data into a plurality of intrinsic mode functions and a residual term through the time series, and differentiating into a high-frequency component and a low-frequency component according to the frequency of the intrinsic mode function; A weight distribution module, used for performing weight partitioning through the intrinsic mode function according to the meteorological data; A combined processing module, used for combining feature extraction with a parallel model according to the high frequency component; A low-frequency prediction module, used to capture nonlinear trends through a gradient boosting regression tree according to the low-frequency components; The prediction reconstruction module is used to add the prediction results of the intrinsic mode functions and recombine them as the prediction value.

[0071] Furthermore, the weight distribution module specifically includes: A construction unit is used to evaluate the relevance of each meteorological data to the electric load by using a Pearson correlation coefficient according to the relevance between the meteorological data and the electric power consumption; An adjustment unit is used to automatically adjust through a neural network according to the correlation and the expert initial weight.

[0072] Embodiment three: This embodiment further provides an electronic device, including a memory and a processor, wherein the memory is used to store one or more computer instructions, wherein the one or more computer instructions implement the method of embodiment 1 when executed by the processor; In practical applications, the processor can be an application specific integrated circuit (ASIC), a digital signal processor (DSP), a digital signal processing device (DSPD), a programmable logic device (PLD), a field programmable gate array (FPGA), a controller, a microcontroller unit (MCU), a microprocessor or other electronic components to execute the methods in the above embodiments.

[0073] The method implemented in this embodiment is as described in Embodiment 1.

[0074] Embodiment 4: This embodiment further provides a computer storage medium, wherein the computer-readable storage medium stores a computer program, and when the computer program is executed by one or more processors, the method of the first embodiment is implemented; Among them, the computer-readable storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk or optical disk.

[0075] The method implemented in this embodiment is as described in Embodiment 1.

[0076] In the several embodiments provided in the embodiments of the present invention, it should be understood that the disclosed system and method can also be implemented in other ways. The system and method embodiments described above are merely illustrative.

[0077] It should be noted that, in this article, the terms "first", "second", etc. in the specification and claims of this application and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. The terms "comprise", "include" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, an element defined by the sentence "comprises a..." does not exclude the presence of other identical elements in the process, method, article or device including the element.

[0078] Although the embodiments disclosed in the present invention are as above, the above contents are only embodiments adopted for facilitating the understanding of the present invention and are not intended to limit the present invention. Any technician in the technical field to which the present invention belongs can make any modifications and changes in the form and details of the implementation without departing from the spirit and scope disclosed in the present invention, but the patent protection scope of the present invention shall still be subject to the scope defined in the attached claims.

Claims

1. A method for power load forecasting based on multimodal feature extraction and decomposition fusion, characterized in that: The method comprises: S1, obtain time series data based on historical charge coincidence and meteorological data, and perform data cleaning; S2, decomposing the time series data into a plurality of intrinsic mode functions and a residual term through the time series, and differentiating the intrinsic mode functions into high-frequency components and low-frequency components according to the frequencies of the intrinsic mode functions; S3, performing weight partitioning by using the intrinsic mode function according to the meteorological data; S4, processing the high frequency components by combining feature extraction with a parallel model; S5, capturing the nonlinear trend through a gradient boosting regression tree according to the low-frequency component; S6, adding the prediction results of the intrinsic mode functions and recombining them as a prediction value.

2. The power load forecasting method based on multimodal feature extraction and decomposition fusion according to claim 1 is characterized in that: The S3 specifically includes: S31, based on the correlation between the meteorological data and the power consumption, evaluating the correlation of each meteorological data to the power load by using the Pearson correlation coefficient.

3. The power load forecasting method based on multimodal feature extraction and decomposition fusion according to claim 2 is characterized in that: The S31 specifically includes: S311, assigning an initial weight to each meteorological factor of the meteorological data according to temperature, humidity and wind speed; S312, according to the Pearson correlation coefficient, the correlation between each meteorological data and the power consumption is analyzed by formula (2); the formula (2) is as follows: (2) In the formula express The average value of express The average value of S313, according to the load data, analyzing the correlation coefficient through each of the meteorological data, wherein the correlation coefficient includes the temperature correlation coefficient ; Humidity correlation coefficient ; Wind speed correlation coefficient ; According to the correlation coefficient, the expert initial weight is adjusted by formula (16); the formula (16) is as follows: (16) in, is the final desired weight.

4. The method for power load forecasting based on multimodal feature extraction and decomposition fusion according to claim 3 is characterized in that: The S3 further includes: S32, automatically adjusting through a neural network according to the correlation and the expert initial weight.

5. The method for power load forecasting based on multimodal feature extraction and decomposition fusion according to claim 4 is characterized in that: The S32 specifically includes: S321, performing model training through the neural network according to the expert initial weights and input data; S322, according to the model training, performing hidden layer processing and feature extraction according to the neural network, and adjusting the expert initial weights through a back propagation algorithm; S323, output a prediction value according to the output layer of the neural network , and the actual power consumption For comparison, the mean square error is used for measurement, and the formula is as follows: (18) in, is the load value predicted by the model; is the actual load value; N is the number of samples; S324, according to the back propagation algorithm, propagate the error forward layer by layer through the chain rule, and adjust the weights and biases of each layer of the neural network.

6. The power load forecasting method based on multimodal feature extraction and decomposition fusion according to claim 1 is characterized in that: The S4 specifically includes: S41, capturing spatiotemporal characteristics through feature extraction according to the high-frequency component, and taking the decomposed high-frequency component and meteorological data as input, performing feature extraction to obtain channel features, spatial features and temporal features; S42, forming a feature vector by feature fusion according to the channel feature, spatial feature and temporal feature.

7. A power load forecasting device based on multimodal feature extraction and decomposition fusion, characterized in that: The device comprises: The preprocessing module is used to obtain time series data based on historical charge coincidence and meteorological data, and perform data cleaning; A sequence decomposition module, used for decomposing the time series data into a plurality of intrinsic mode functions and a residual term through the time series, and differentiating into a high-frequency component and a low-frequency component according to the frequency of the intrinsic mode function; A weight distribution module, used for performing weight partitioning through the intrinsic mode function according to the meteorological data; A combined processing module, used for combining feature extraction with a parallel model according to the high frequency component; A low-frequency prediction module, used to capture nonlinear trends through a gradient boosting regression tree according to the low-frequency components; The prediction reconstruction module is used to add the prediction results of the intrinsic mode functions and recombine them as the prediction value.

8. The power load forecasting device based on multimodal feature extraction and decomposition fusion according to claim 7 is characterized in that: The weight distribution module specifically includes: A construction unit is used to evaluate the relevance of each meteorological data to the electric load by using a Pearson correlation coefficient according to the relevance between the meteorological data and the electric power consumption; An adjustment unit is used to automatically adjust through a neural network according to the correlation and the expert initial weight.

9. An electronic device, characterized in that: It comprises a memory and a processor, wherein the memory is used to store one or more computer instructions, wherein the one or more computer instructions, when executed by the processor, implement the power load forecasting method based on multimodal feature extraction and decomposition fusion as described in any one of claims 1 to 6 above.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, it is used to implement the power load forecasting method based on multimodal feature extraction and decomposition fusion as described in any one of claims 1 to 6 above.

Images