A power load decomposition and identification method and system based on TCN and BiGRU

Abstract

The application relates to a TCN and BiGRU-based power load decomposition and identification method and system, which comprises the following steps: acquiring a data set of total household power and power of each electrical appliance, and performing data preprocessing on the data set to establish a training data set and a test data set; a TCN_BiGRU network model comprising a TCN network, a BiGRU network and a GAM attention mechanism is constructed; a power sequence is input into the TCN network for preliminary feature extraction, long-time sequence dependency information is extracted, and shallow features of the power sequence are obtained; the shallow features are used as input features of the BiGRU network to extract bidirectional deep sequence features of the shallow features; the bidirectional deep sequence features are used as input sequence information of the GAM attention mechanism to output a predicted sequence midpoint; the TCN_BiGRU network model is trained and verified through the training data set and the test data set; and the trained TCN_BiGRU network model is used for power load decomposition and identification. The method and system are beneficial to the extraction of long-time dependency information and the improvement of power load decomposition and identification precision.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of non-intrusive power load decomposition, and particularly relates to a power load decomposition and identification method and system based on TCN and BiGRU. BACKGROUND

[0002] The load monitoring method can help users understand their own power consumption behavior and regulate their own power consumption habits, and can also help power grid companies to realize scientific power dispatching. The general load monitoring method is divided into non-intrusive load monitoring and intrusive load monitoring. The intrusive power load monitoring needs to use multiple sensors to realize monitoring, and the non-intrusive power load monitoring can realize the target only through the smart meter at the incoming line end. It has the characteristics of low cost, good privacy, is conducive to wide promotion and has high user acceptance.

[0003] In recent years, various non-intrusive power load decomposition methods have been proposed. Han Xiao proposed using high-frequency V-I curve for non-intrusive power load decomposition, which achieved good results. However, the high-frequency acquisition cost is high, which is not conducive to wide promotion. Therefore, most researches are based on low-frequency data. Machine learning algorithms are widely used in the field of non-intrusive power load decomposition, including decision tree (DT), support vector machine (SVM), k-nearest neighbor, and other machine learning methods. The advantages are that the required data set is small, the training time is short, but the generalization performance is weak and the decomposition accuracy is low. In recent years, deep learning methods have gradually replaced machine learning methods and become the mainstream of research. Kelly et al. first used denoising autoencoder (DAE), LSTM and other models for low-frequency power load decomposition, and proved that the accuracy is higher than that of machine learning algorithms. Zhang C et al. applied a fully convolutional neural network to the power load decomposition task and achieved good accuracy. In addition, Seq2Seq model, transformer, CNN_LSTM and other models are also used in this task. These models have improved the decomposition accuracy and feature extraction ability to some extent, and can effectively identify the power curves of electrical appliances in the total power curve. However, these models all face some problems such as insufficient feature extraction, inability to consider global features, and inability to identify long-term dependencies in power sequences. SUMMARY

[0004] The present application relates to the technical field of non-intrusive power load decomposition, and particularly relates to a power load decomposition and identification method and system based on TCN and BiGRU.

[0005] In order to achieve the above-mentioned purpose, the technical scheme adopted by the present application is as follows: a power load decomposition and identification method based on TCN and BiGRU, comprising the following steps:

[0006] Step S1: Obtain the dataset of total household power consumption and power consumption of each appliance, and perform data preprocessing on the dataset to establish training and test datasets.

[0007] Step S2: Construct a TCN_BiGRU network model including a TCN network, a BiGRU network, and a GAM attention mechanism; input the power sequence into the TCN network for preliminary feature extraction, extract long-term dependency information of the sequence, and obtain shallow features of the power sequence; use the obtained shallow features as input features of the BiGRU network to extract bidirectional deep sequence features from the shallow features; use the bidirectional deep sequence features as input sequence information of the GAM attention mechanism, consider global sequence features while retaining key information, reducing information loss and amplifying global dimension features, and outputting the predicted sequence midpoint; train and validate the TCN_BiGRU network model using training and testing datasets to obtain the trained TCN_BiGRU network model.

[0008] Step S3: Decompose and identify power load using the trained TCN_BiGRU network model.

[0009] Furthermore, in step S1, the data preprocessing method for the dataset is as follows:

[0010] Step S11: Clear outliers from the total power sequence and fill the missing values ​​with the value of the previous point in the total power sequence;

[0011] Step S12: For the total power sequence [X1, X2, X3, ..., X...] of the n sampling points after removing outliers... n The power sequence is divided into fixed-length power sequences using a sliding window, and the midpoint of the power sequence of the applied electrical appliance is taken as the label value.

[0012] Step S13: Divide the segmented power sequence and labels into training dataset and test dataset according to a set ratio.

[0013] Furthermore, in step S2, the TCN network is implemented as follows:

[0014] Temporal convolution is decomposed into causal convolution, dilated convolution, and residual networks; the time series is in x t The formula for calculating causal convolution at a given point is as follows:

[0015]

[0016] Where F is the filter sequence [f1,f2,f3,...,f K ], f kLet X be the value of the k-th filter in the filter sequence F, and let X be the time series [x1, x2, x3, ..., xk]. T ], x t-k Let t be the k-th time point in the time series X. As can be seen from the above formula, the causal convolution calculation only uses the sequence values ​​before time t and is independent of the sequence values ​​after time t. Therefore, it has causal properties.

[0017] Time series in x t The formula for calculating dilated convolution at a given point is as follows:

[0018]

[0019] Where F is the filter sequence [f1,f2,f3,...,f K ], f k Let X be the value of the k-th filter in the filter sequence F, and let X be the time series [x1, x2, x3, ..., xk]. T ], x t-(K-k)d Let d be the (t-(Kk)d)th time point in the time series X, where d is the inflation factor;

[0020] Residual connections are introduced to transfer information across layers, thus mitigating the vanishing and exploding gradient problems. The formula for calculating residual connections is as follows:

[0021] y = W 1x1 ·x+b

[0022] Where b is the bias term, W 1×1 is a 1×1 convolution kernel, and x is the feature map calculated by the temporal convolution;

[0023] After the input time series data is processed by causal convolution and dilated convolution to extract features, a residual network is used to link and transmit information across layers, and finally the long-term dependence information of the sequence is extracted, and shallow feature information is output.

[0024] Furthermore, in step S2, the BiGRU network is implemented as follows:

[0025] A BiGRU network is implemented using unidirectional GRU overlap in opposite directions; in the GRU, the output of the update gate is represented as follows:

[0026] z t =σ(W r ·[h t-1 ,x t ]+b z )

[0027] The output of the reset gate is represented as follows:

[0028] r t =σ(W r ·[ht-1 , t ]+b r )

[0029] The candidate hidden state is calculated by the reset gate, and the output of the candidate hidden state is represented as follows:

[0030]

[0031] The updated hidden state is calculated by the candidate state and the update gate, and the output of the updated hidden state is represented as follows:

[0032]

[0033] wherein σ represents a Sigmoid activation function, tanh represents a hyperbolic tangent activation function, W r , W, and W are weight matrices, x t is a shallow feature extracted by the TCN, b r , b z , and b h are biases, and h t-1 is a hidden state feature of the previous time step;

[0034] The BiGRU network is composed of two opposite GRUs, and bidirectional deep time sequence features are extracted.

[0035] Further, in step S2, the GAM attention mechanism includes a channel attention module and a spatial attention module, and the implementation method is as follows:

[0036] Step S41: input the load feature map F1, F1∈R H×W×C , wherein C is the channel number of the input load feature map, H and W are the height and width of the input load feature map, the input load feature map F1 is redistributed by the channel attention module to generate a feature weight M C (F1), and the feature map F2 is obtained by point multiplication of M S (F1) and the input load feature map F1, and is output, and the specific formula is as follows:

[0037]

[0038] Step S42: input the output feature map F2 into the spatial attention module to generate an attention weight M S (F2), multiply M S (F2) weight and the feature map F2 to obtain a load feature F3 with the same feature dimension as the original feature; and the specific formula is as follows:

[0039]

[0040] Step S43: two-layer full connection layers are performed on the sequence features extracted by the GAM attention mechanism to realize dimension change, and then a sequence midpoint prediction value is output, wherein the full connection layer is represented as follows:

[0041] z1=W1*x1+b1

[0042] z2=W2*z1+b2

[0043] wherein x1 is a feature map output by the GAM attention mechanism, W1 and W2 are weight matrices, z1 is an output vector of the first layer full connection layer, z2 is an output vector of the second layer full connection layer, and b1 and b2 are biases.

[0044] Further, in step S2, the implementation method of training and verifying the TCN_BiGRU network model through the training data set and the test data set is as follows:

[0045] Adam training parameters are set, the model is trained through the training data set in batches, the network weight of the best batch is taken as the model training result, the result is verified in the test data set, F1 and MAE are taken as the model evaluation indexes, the average absolute error between the predicted appliance power prediction value and the true value is calculated to evaluate the model decomposition performance, and F1 is taken to evaluate the model recognition performance of the appliance opening and closing.

[0046] The application also provides a power load decomposition and recognition system based on TCN and BiGRU, which comprises a memory, a processor and computer program instructions stored in the memory and capable of being executed by the processor, and when the processor executes the computer program instructions, the above-mentioned method steps can be realized.

[0047] Compared with the prior art, the application has the following beneficial effects: the application provides a power load decomposition and recognition method and system based on TCN and BiGRU, the method inputs the power sequence into the TCN network to extract long-time dependence in the sequence, obtains shallow sequence features, inputs the shallow features into the BiGRU network to extract deep bidirectional time sequence features, then obtains bidirectional time sequence feature information through the GAM attention mechanism, considers global sequence features while retaining key information, reduces information loss and releases global dimension features, and finally outputs sequence midpoint information through dimension change of the full connection layer. Through the method and system provided by the application, long-time dependence information can be better extracted, the power load decomposition and recognition precision can be improved, and the model training time can be shortened. BRIEF DESCRIPTION OF DRAWINGS

[0048] Figure 1 The method of the embodiment of the application realizes a flowchart;

[0049] Figure 2This is an overall structure diagram of the TCN_BiGRU network model in an embodiment of the present invention;

[0050] Figure 3 This is a schematic diagram of sliding window data preprocessing in an embodiment of the present invention;

[0051] Figure 4 This is a diagram of the temporal convolution structure in an embodiment of the present invention;

[0052] Figure 5 This is a diagram of the BiGRU network structure in an embodiment of the present invention;

[0053] Figure 6 This is a structural diagram of the GAM attention mechanism in an embodiment of the present invention. Detailed Implementation

[0054] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0055] It should be noted that the following detailed descriptions are exemplary and intended to provide further explanation of this application. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains.

[0056] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments according to this application. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0057] like Figure 1 As shown, this embodiment provides a method for power load decomposition and identification based on TCN and BiGRU, including the following steps:

[0058] Step S1: Obtain the dataset of total household power consumption and the power consumption of each appliance, and perform data preprocessing on the dataset to establish training and test datasets.

[0059] Step S2: Construct the TCN_BiGRU network model, such as... Figure 2As shown, the model mainly includes a TCN network, a BiGRU network, and a GAM attention mechanism. The power sequence is input into the TCN network for preliminary feature extraction, sequence long-time dependency information is extracted, and shallow features of the power sequence are obtained. The obtained shallow features are taken as input features of the BiGRU network to extract bidirectional deep sequence features of the shallow features. The bidirectional deep sequence features are taken as input sequence information of the GAM attention mechanism to consider global sequence features while retaining key information, reduce information loss, and put all global dimension features. A middle point of a predicted sequence is output. The TCN_BiGRU network model is trained and verified through a training data set and a test data set, and a trained TCN_BiGRU network model is obtained.

[0060] Step S3, power load decomposition and identification are performed through the trained TCN_BiGRU network model.

[0061] As shown in Figure 3 , the implementation method of data preprocessing of the data set is as follows:

[0062] Step S11, the abnormal values in the total power sequence are cleared, and the value of the previous point of the total power sequence is taken to fill the missing values.

[0063] In this embodiment, the specific method of cleaning abnormal values is to detect abnormal values in the total power sequence, remove the power point of the abnormal values, and fill the abnormal values using the value to the left of the abnormal values.

[0064] Step S12, the total power sequence [X1, X2, X3,..., X n ] of n sampling points after the abnormal values are cleared is segmented into fixed-length power sequences by using a sliding window, the window size is taken as 127, and the midpoint of the power sequence of the appliance is taken as a label value.

[0065] In this embodiment, the specific method of segmenting data by using a sliding window is that, since the length of the total power curve is indefinite, the length of the time series data required by the deep learning network is fixed, and therefore the total power data needs to be segmented by using a sliding window. The process of processing power data by using a sliding window is as shown in Figure 3 . The size of the sliding window is the required sequence size, the sequence data [X1, X2, X3,..., X n ] in the sliding window is segmented each time, the sliding window is moved by a fixed step n after the sequence is segmented, new time series data is segmented, and so on until the window slides to the end of the sequence.

[0066] Step S13, the segmented power sequence and the label are divided into a training data set and a test data set according to a set proportion.

[0067] In the embodiment, the specific division method of the training set and the test set is: 5 commonly used electric appliance power curves and a total power curve are selected to form the training set and the test set, and in the embodiment, an air conditioner, a kettle, a dishwasher, a washing machine and a microwave oven are selected as labels. In this paper, the Seq2point output method is used, and the sequence point corresponding to the total power in the sliding window of the electric appliance sequence is taken as the label output. A total of five groups of electric appliance data sets are constructed, and finally the data sets are normalized and divided into 7:3 training set and test set data.

[0068] In the embodiment, the structure of the TCN network (time convolution network) is as shown in the figure. Figure 4 The value of each layer of time convolution is only related to the corresponding position of the previous layer and the previous convolution unit. The time convolution network combines the convolution network of dilated convolution and causal convolution. For causal convolution, the output of each time step is only related to the input of the corresponding time step of the previous layer and the input of the previous time step, and is not dependent on future data, so it has causality, and the disadvantage is that the receptive field is small, while the dilated convolution has a larger receptive field and can sense longer time series data. The time convolution combines the advantages of the two kinds of convolution, has causality and a larger receptive field, can better extract the long-time dependence of time series data and extract deeper feature relationships.

[0069] The time convolution is divided into causal convolution, dilated convolution and residual network; the causal convolution calculation formula of the time series at x t is as follows:

[0070]

[0071] Among them, F is the filter sequence [f1, f2, f3,..., f K ], f k is the value of the kth filter in the filter sequence F, X is the time series [x1, x2, x3,..., x T ], x t-k is the t-kth time point in the time series X, and according to the above formula, the causal convolution calculation only uses the sequence values before t, and is independent of the sequence values after t, so it has the causal characteristic.

[0072] The dilated convolution calculation formula of the time series at x t is as follows:

[0073]

[0074] Among them, F is the filter sequence [f1, f2, f3,..., f K ], f k is the value of the kth filter in the filter sequence F, X is the time series [x1, x2, x3,..., xT ],x t-(K-k)d is the t-(K-k)dth time point in the time series X, d is the dilation factor, in the dilated convolution, the filter with the same width has a larger receptive field, which can extract longer time series features.

[0075] The time convolution network is deep, so the residual connection is introduced to pass information across layers to alleviate the gradient vanishing and gradient explosion problems; the residual connection calculation formula is as follows:

[0076] y=W 1x1 ·x+b

[0077] Where b is the bias term, W 1×1 is a 1x1 convolution kernel, x is the feature map calculated by time convolution, and the residual connection is used to alleviate gradient explosion and gradient vanishing.

[0078] After the input time series data is extracted by the causal convolution and the dilated convolution, the residual network link is used to pass information across layers, and finally the long-time dependence information of the sequence is extracted, and the shallow feature information is output.

[0079] In this embodiment, the structure of the BiGRU network is as shown in Figure 5 The BiGRU network is realized by using opposite unidirectional GRU overlap, and the output vectors of two layers of opposite GRU are fused to output features. A single-layer GRU network can only extract forward time series features, and some reverse time series relationships are easily ignored. Using BiGRU can effectively avoid this situation. GRU is a modified network unit of LSTM, which introduces update gate and reset gate units, and has smaller calculation amount than LSTM network.

[0080] In the GRU, the output of the update gate is represented as follows:

[0081] z t =σ(W r ·[h t-1 ,x t ]+b z )

[0082] The output of the reset gate is represented as follows:

[0083] r t =σ(W r ·[h t-1 ,x t ]+b r )

[0084] The candidate hidden state is calculated by the reset gate, and the output of the candidate hidden state is represented as follows:

[0085]

[0086] The update hidden state is calculated by the candidate state and the update gate, and the output of the update hidden state is represented as follows:

[0087]

[0088] wherein σ represents a Sigmoid activation function, tanh represents a hyperbolic tangent activation function, W r , W, and W are weight matrices, x t is a shallow feature extracted by the TCN, b r , b z , and b h are biases, h t-1 is a hidden state feature of the previous time step.

[0089] The GRU unit is fast in calculation and good in performance, and the BiGRU network is composed of two opposite GRUs. The shallow features of the TCN are input into the BiGRU network to extract bidirectional sequence features and obtain deepened sequence features.

[0090] In this embodiment, the structure of the BiGRU network is as shown in Figure 6 The GAM attention mechanism includes a channel attention module and a spatial attention module, and the implementation method is as follows:

[0091] Step S41: input the load feature map F1, F1∈R H×W×C , wherein C is the channel number of the input load feature map, H and W are the height and width of the input load feature map, the input load feature map F1 is re-distributed by the channel attention module to generate a feature weight M C (F1), and the feature map F2 is obtained by point multiplication of M S (F1) and the input load feature map F1, and is output, and the specific formula is as follows:

[0092]

[0093] Step S42: input the output feature map F2 into the spatial attention module to generate an attention weight M S (F2), multiply M S (F2) with the feature map F2 to obtain a load feature F3 with the same feature dimension as the original feature; and the specific formula is as follows:

[0094]

[0095] Step S43: two fully connected layers are performed on the sequence features re-extracted by the GAM attention mechanism to realize dimension change, and then a point prediction value in the sequence is output, and the fully connected layer is represented as follows:

[0096] z1=W1·x1+b1

[0097] z2 = W2 * z1 + b2

[0098] Wherein, x1 is the feature map output by the GAM attention mechanism, W1 and W2 are weight matrices, z1 is the output vector of the first full connection layer, z2 is the output vector of the second full connection layer, b1 and b2 are biases. The final output of z2 is the predicted sequence midpoint.

[0099] The implementation method of training and verifying the TCN_BiGRU network model through the training data set and the test data set is as follows:

[0100] The Adam training parameters are set, the model is trained through the training data set in batches, the network weight of the best batch is taken as the model training result, the result is verified in the test data set, F1 and MAE are taken as the model evaluation indexes, the average absolute error between the predicted value and the true value of the predicted power of the electric appliance is calculated, so as to evaluate the decomposition performance of the model, and F1 is taken to evaluate the recognition performance of the model to the opening and closing of the electric appliance. In this embodiment, the Adam parameters lr is set to 0.001, beta1 is set to 0.9, and beta2 is set to 0.999.

[0101] The embodiment also provides a power load decomposition and recognition system based on TCN and BiGRU, which comprises a memory, a processor and computer program instructions stored in the memory and capable of being executed by the processor, and when the processor executes the computer program instructions, the above-mentioned method steps can be realized.

[0102] Those skilled in the art should understand that the embodiments of the present application can be provided as methods, systems or computer program products. Therefore, the present application can adopt a complete hardware embodiment, a complete software embodiment or an embodiment combining software and hardware aspects. Moreover, the present application can adopt the form of a computer program product implemented on one or more computer usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer usable program codes.

[0103] The present application is described with reference to flowcharts and / or block diagrams according to the methods, devices (systems) and computer program products of the embodiments of the present application. It should be understood that each flow and / or block in the flowcharts and / or block diagrams, and the combination of the flows and / or blocks in the flowcharts and / or block diagrams can be realized by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device produce a machine that implements the functions described in the flowcharts and / or block diagrams. Figure 1 one flow or multiple flows and / or blocks Figure 1means for performing the function specified by the block or blocks.

[0104] These computer program instructions can also be stored in a computer readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer readable memory produce an article of manufacture including instructions which implement the flow Figure 1 one or more flowcharts and / or blocks Figure 1 means for performing the function specified by the block or blocks.

[0105] These computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the flow Figure 1 one or more flowcharts and / or blocks Figure 1 means for performing the function specified by the block or blocks.

[0106] The above descriptions are only preferred embodiments of the present application, and are not intended to limit the present application to other forms described. Any person skilled in the art may make modifications or improvements to the above-described technical content, or make equivalent changes to equivalent embodiments. However, any simple modification, equivalent change and modification of the above embodiments, which does not deviate from the technical content of the present application, and which is based on the technical essence of the present application, still falls within the protection scope of the present application.

Claims

1. A TCN and BiGRU-based power load decomposition and identification method, characterized in that, The method comprises the following steps: Step S1, obtaining a data set of total power consumption of a household and power consumption of each electrical appliance, and performing data preprocessing on the data set to establish a training data set and a test data set; Step S2, constructing a TCN_BiGRU network model comprising a TCN network, a BiGRU network and a GAM attention mechanism; inputting a power sequence into the TCN network to perform preliminary feature extraction, extract long-time sequence dependence information, and obtain shallow features of the power sequence; The obtained shallow features are taken as input features of the BiGRU network to extract bidirectional deep sequence features of the shallow features; the bidirectional deep sequence features are taken as input sequence information of the GAM attention mechanism to consider global sequence features while retaining key information, reduce information loss and release global dimension features, and output a predicted sequence midpoint; the TCN_BiGRU network model is trained and verified through the training data set and the test data set to obtain a trained TCN_BiGRU network model; Step S3, performing power load decomposition and identification through the trained TCN_BiGRU network model.

2. The TCN and BiGRU-based power load decomposition and identification method according to claim 1, characterized in that, In step S1, the implementation method of data preprocessing on the data set is as follows: Step S11, removing abnormal values in the total power sequence, and filling the missing values with the value of the previous point of the total power sequence; Step S12: For the total power sequence [X1, X2, X3, ..., X] of the n sampling points after removing outliers... n The power sequence is divided into fixed-length power sequences using a sliding window, and the midpoint of the power sequence of the applied electrical appliance is taken as the label value. Step S13, dividing the segmented power sequence and the label into a training data set and a test data set according to a set proportion.

3. The TCN and BiGRU-based power load decomposition and identification method of claim 1, wherein, In step S2, the implementation method of the TCN network is as follows: Splitting the temporal convolution into a causal convolution, a dilated convolution, and a residual network; the temporal series at x t The causal convolution computation formula at x is as follows: Where F is the filter sequence [f1, f2, f3, ..., f K ], f k Let X be the value of the k-th filter in the filter sequence F, and let X be the time series [x1, x2, x3, ..., x...]. T ], x t-k Let t be the k-th time point in the time series X. As can be seen from the above formula, the causal convolution calculation only uses the sequence values ​​before time t and is independent of the sequence values ​​after time t. Therefore, it has causal properties. The expansion convolution calculation formula at time series x t is as follows: Where F is the filter sequence [f1, f2, f3, ..., f K ], f k Let X be the value of the k-th filter in the filter sequence F, and let X be the time series [x1, x2, x3, ..., x...]. T ], x t-(K-k)d Let d be the (t-(Kk)d)th time point in the time series X, where d is the inflation factor; Residual connection is introduced to transfer information across layers to alleviate the problems of gradient disappearance and gradient explosion; the calculation formula of the residual connection is as follows: where b is a bias term, W 1×1 is a 1x1 convolution kernel, and x is a feature map calculated by time convolution. After the input time series data extracts features through causal convolution and dilated convolution, the residual network link is adopted to transfer information across layers, and finally the long-time sequence dependence information is extracted, and the shallow feature information is output.

4. The TCN and BiGRU-based power load decomposition and identification method of claim 1, wherein, In step S2, the implementation method of the BiGRU network is as follows: The BiGRU network is realized by using opposite unidirectional GRUs; in the GRU, the output of the update gate is represented as follows: The output of the reset gate is represented as follows: The candidate hidden state is calculated through the reset gate, and the output of the candidate hidden state is represented as follows: The update hidden state is calculated from the candidate state and the update gate, and the output of the update hidden state is represented as follows: wherein, denotes a Sigmoid activation function, tanh denotes a hyperbolic tangent activation function, W r , W h is a weight matrix, x t is a shallow feature extracted by the TCN, b r , b z and b h is a bias, h t-1 is a hidden state feature of the previous time step; The BiGRU network is composed of two opposite GRUs, and the bidirectional deep time sequence features are extracted.

5. The TCN and BiGRU-based power load decomposition and identification method of claim 1, wherein, In step S2, the GAM attention mechanism comprises a channel attention module and a spatial attention module, and the implementation method is as follows: Step S41: input the load feature map , , where C is the number of channels of the input load feature map, H and W are the height and width of the input load feature map, the input load feature map is redistributed by the channel attention module to generate a feature weight , and the input load feature map is multiplied by the feature weight to obtain a feature map F2 and output, and the specific formula is as follows: Step S42: input the output feature map F2 into the spatial attention module to generate attention weight M S (F2), multiply M S (F2) weight with the feature map F2 to obtain a load feature F3 with the same original feature dimension; the specific formula is as follows: Step S43: two fully connected layers are used to realize dimension change on the sequence features reextracted through the GAM attention mechanism, and then a sequence midpoint prediction value is output, wherein the fully connected layer is represented as follows: Wherein, x1 is the feature map output by the GAM attention mechanism, W1 and W2 are weight matrices, z1 is the output vector of the first fully connected layer, z2 is the output vector of the second fully connected layer, b1 and b2 are biases.

6. The TCN and BiGRU-based power load decomposition and identification method of claim 1, wherein, In step S2, the implementation method of training and verifying the TCN_BiGRU network model through the training data set and the test data set is as follows: Adam training parameters are set, the model is trained in batches through the training data set, the network weight of the best batch is taken as the model training result; the result is verified in the test data set, F1 and MAE are taken as the model evaluation indexes, the average absolute error of the predicted value and the true value of the predicted power of the electric appliance is calculated, so as to evaluate the decomposition performance of the model, and F1 is taken to evaluate the recognition performance of the model on the opening and closing of the electric appliance.

7. A TCN and BiGRU-based power load decomposition and identification system, characterized in that, The computer program instructions stored in the memory and capable of being executed by the processor can implement the method steps of any one of claims 1-6 when the processor executes the computer program instructions.

Images