A non-invasive load monitoring model aggregation method and system

Abstract

The present invention discloses a non-intrusive load monitoring model aggregation method and system. The method includes: obtaining the active power of the user bus and each load; pairing the active power of the user bus and the active power of each load to form a data set and performing preprocessing; initially training three preset base models, and the three initially trained base models decompose the active power of each load to obtain a decomposition result; using a pre-established perceptron aggregation model to aggregate the decomposition result and optimize the pre-established perceptron aggregation model to obtain preliminary aggregation data of the active power of each load; optimizing a pre-established DAE model based on the decomposition result and the preliminary aggregation data, and the optimized DAE model is the aggregation model of the non-intrusive load monitoring model, which is used to output the data of the active power of each load. The present invention can aggregate the decomposition results of different loads to obtain a decomposition result superior to that of a single model.

Description

Technical Field

[0001] The present invention relates to a non-invasive load monitoring model aggregation method and system, belonging to the technical field of non-invasive electricity load monitoring. Background Art

[0002] At first, when measuring and obtaining data for household appliances, many were monitored in an invasive manner, that is, sensors with advanced communication functions were installed between each appliance and the distribution box. Although this method has high measurement accuracy, the installation or transformation of hardware will bring inconvenience to residents' lives, and the structure is often complex, costly, and the later maintenance is also rather cumbersome. Therefore, a non-invasive load monitoring structure was proposed to fill this defect, which is equivalent to a one-to-many relationship. Take a single household as an example. If an intelligent electricity meter is installed at the entrance of the household, the electricity consumption information of the user can be collected, and the electricity consumption of each device can be obtained, thereby greatly reducing the hardware cost, facilitating installation, and being suitable for household load monitoring.

[0003] Regarding the non-invasive load identification algorithm, current non-invasive power load intelligent identification systems generally aggregate based on a single model, such as common models like deep learning models. A single model often has better identification effects only for certain types of electrical loads, and its universality and generality need to be improved. There have been many deep learning algorithms with good load monitoring effects applied to non-invasive load monitoring. For example, the Seq2Point model, DAE model, and attention-based residual network model, etc. Each model has different accuracies in decomposing power for different load devices, and currently, there is a lack of a method for aggregating the decomposition results of each model. In the process of load decomposition using existing models, a certain electrical appliance is recognized well, but the recognition effect for other electrical appliances is poor. The present invention aims to achieve comprehensive optimization to ensure the best decomposition effect for each load. Summary of the Invention

[0004] The purpose of the present invention is to overcome the deficiencies in the prior art and provide a non-invasive load monitoring model aggregation method and system that can aggregate the decomposition results of different loads to obtain decomposition results superior to those of a single model.

[0005] To achieve the above object, the present invention is implemented by the following technical solutions:

[0006] In the first aspect, the present invention provides a non-invasive load monitoring model aggregation method, including:

[0007] Obtain the active power of the user bus and each load;

[0008] Pair the active power of the user bus and the active power of each load to form a data set, and process the data set;

[0009] Perform the initial training on three preset base models based on the processed dataset, and use the three base models completed in the initial training to decompose the active power of each load to obtain the decomposition results;

[0010] Use the pre-established perceptron aggregation model to aggregate the decomposition results and optimize the parameters of the pre-established perceptron aggregation model. Based on the trained perceptron aggregation model, obtain the preliminary aggregated data of the active power of each load;

[0011] Optimize the pre-established DAE model based on the decomposition results and the preliminary aggregated data. The optimized DAE model is the aggregation model of the non-intrusive load monitoring model, which is used to output the data of the active power of each load.

[0012] Combined with the first aspect, further, the obtaining of the user bus and the active power of each load includes: using the active power monitoring device to extract the active power data of each load and load cluster of the user at a frequency of times / minute, arranging them in chronological order to obtain the active power of the user bus and the active power of each load.

[0013] Combined with the first aspect, further, the composition of the dataset includes:

[0014] Obtaining the active power data of a specified length by using a sliding window includes: setting the length of the time series data intercepted by the sliding window to be T, and sliding and intercepting the active power of the user bus and the active power of each load obtained in chronological order step by step with a step size of 1. The total length of the data is N,

[0015] Then the data of the active power of the user bus is:

[0016] The data of the active power of the i-th load is:

[0017] The pairing of the active power of the user bus and the active power of each load to form a dataset includes:

[0018] Let the data of the active power of the user bus where x n is a vector of T, n = 1, 2,..., N - T + 1,

[0019] The data of the active power of the i-th load where is a vector of T, n = 1, 2,..., N - T + 1;

[0020] Then is the dataset of the pairing of the active power of the user bus and the active power of each load.

[0021] In combination with the first aspect, further, the processing of the data set includes:

[0022] Perform normalization processing on the data, including: extracting the maximum value of the active power of the user bus, and dividing the active power of the user bus and the active power of each load by the maximum value of the active power of the user bus respectively.

[0023] In combination with the first aspect, further, the initial training of the three preset basic models includes:

[0024] Use the processed data set of the i-th load to train the preset Seq2Point model, and update the model parameters using the reverse optimization algorithm to obtain the trained Seq2Point model;

[0025] Use the processed data set of the i-th load to train the preset DAE model, and update the model parameters using the stochastic gradient descent algorithm to obtain the trained DAE model;

[0026] Use the processed data set of the i-th load to train the preset Attention and residual network model, update the Attention parameters using the BP algorithm to obtain the trained Attention, and based on the trained Attention, update the model parameters using the stochastic gradient descent algorithm to obtain the trained residual network model;

[0027] Use the three initially trained basic models to decompose the active power of each load in the processed data set for x n The decomposition results of the three trained basic models for the i-th load are respectively:

[0028] The decomposition result of the trained Seq2Point model:

[0029] The decomposition result of the trained DAE model:

[0030] The decomposition result of the trained residual network model:

[0031] In combination with the first aspect, optionally, the preset Seq2Point model includes:

[0032] Input layer: Input The T sampling points into the preset Seq2Point model, and the input data tensor is (T, 1);

[0033] Forward GRU layer: Discover the internal change rules in the data of the input layer. The hidden layer size is 2T. The input data tensor of the second forward GRU layer is (T, 2T), and the output data tensor is (4T, 1);

[0034] Fully connected layer: Map high-dimensional data to low-dimensional data. The input data tensor is (4T, 1), and the output data tensor is (2T, 1);

[0035] Output layer: Output the normalized power value of the i-th load. The input data tensor is (2T, 1), and the output data tensor is (1, 1).

[0036] Combined with the first aspect, optionally, the training of the preset Seq2Point model includes:

[0037] The activation function used in each unit is the Relu function, expressed as:

[0038]

[0039] The calculation formula of the GRU layer, expressed as:

[0040]

[0041] In formula (2), r t represents the reset gate, W r represents the weight of the reset gate, σ represents the sigmoid function, h t-1 represents the state passed down from the previous neuron, z t represents the state passed down from the previous neuron, W z represents the weight of the update gate, represents the current neuron state, represents the weight value;

[0042] Use the mean squared error as the loss function, and the expression is:

[0043]

[0044] In formula (3), MSE represents the loss function, m represents the number of samples of the data for training, represents the active power predicted by the preset Seq2Point model;

[0045] Adopt the reverse optimization algorithm to update the model parameters to make the model converge and obtain the trained Seq2Point model.

[0046] Combined with the first aspect, optionally, the preset DAE model includes:

[0047] Encoder: Used to train data and generate a set of feature maps, successively including a convolutional layer, a linear activation function, a max pooling layer, additional convolutional and pooling layers, a fully connected layer, and an encoder network with the ReLU activation function turned off;

[0048] Decoder: Successively including a ReLU activation function, a fully connected layer, additional convolutional and pooling layers, an upsampling layer, a linear activation function, and a convolutional layer.

[0049] Combined with the first aspect, optionally, the preset Attention and residual network model includes:

[0050] Input layer: Used to input the feature x of each sample n ;

[0051] Embedding layer: Used to generate a low-dimensional dense vector for feature extraction. The transformed low-dimensional dense vector is:

[0052] h = WX + b (4)

[0053] In formula (4), h represents the output of the hidden layer of the preset Attention and residual network model, W represents the transformation matrix, b represents the coefficient matrix, and the parameters of W and b are updated based on the BP algorithm during the model training process;

[0054] Encoder, decoder: Used to splice the vectors output by the forward GRU and the backward GRU;

[0055] Output layer: Used to output the result vector.

[0056] Combined with the first aspect, optionally, the training of the preset Attention includes:

[0057] According to the attention mechanism, using the hidden state sequence [h 0 , h 1 ,..., h n-1 , convert the fixed semantic vector c into a dynamic variable semantic vector c t , define the parameter vectors Q, K, V, expressed as follows:

[0058]

[0059] In formula (5), the parameter vectors Q, K, V are updated based on the BP algorithm during the model training process, then:

[0060] Q = K = V = h = [h 0 , h 1 ,..., h n-1 (6)

[0061] Let the function between the residual connections of the residual network be f(x), then the output of the residual network is g(X) = X + f(X);

[0062] Using cross-entropy as the loss function and the stochastic gradient descent algorithm to update the model parameters, a trained residual network model is obtained.

[0063] Combined with the first aspect, further, the obtaining of the preliminary aggregation data of each load active power includes:

[0064] Form the decomposition results into a decomposition data set n = 1, 2, …, N - T + 1, where represents the feature, represents the target value;

[0065] Use the pre-established perceptron aggregation model to aggregate the features to obtain:

[0066]

[0067] Use the decomposition data set to train the pre-established perceptron aggregation model, select the mean squared error as the loss function, and use the stochastic gradient descent algorithm to update the parameters of the pre-established perceptron aggregation model to obtain a trained perceptron aggregation model;

[0068] Based on the trained perceptron aggregation model, obtain the preliminary aggregation data of each load active power n = 1, 2, …, N - T + 1.

[0069] Combined with the first aspect, further, the optimization of the pre-established DAE model includes:

[0070] The decomposition results and the preliminary aggregation data are composed into a training data set;

[0071] Use the training data set to train the pre-established DAE model, select the mean squared error as the loss function, and use the stochastic gradient descent algorithm to update the parameters of the pre-established DAE model to obtain a trained DAE model.

[0072] Combined with the first aspect, optionally, the model structure of the pre-established DAE model has 7 layers, which are in turn:

[0073] The length of the first layer input layer is determined by the input sequence length;

[0074] The second layer convolutional layer processes the input signal, extracts features to form a feature map, has 8 convolutional kernels of length 4, and uses a linear activation function;

[0075] The feature map is expanded through the third layer and then compressed in size through two fully connected layers, namely the fourth and fifth layers, to reduce the dimension of the sequence in order to obtain a compact form of the input sequence, retaining the valid information while removing the noise information therein;

[0076] Taking the fifth layer as the midpoint, the sixth and seventh layers are symmetric with the second and fourth layers. The seventh layer is a convolutional layer. The sixth and seventh layers are used to decode the data in the compact form output by the first five layers to obtain an output sequence in the same form as the input sequence.

[0077] In a second aspect, the present invention provides a non-intrusive electrical load monitoring multi-model aggregation system, including:

[0078] An acquisition module: used to acquire the active power of the user bus and each load;

[0079] A data preprocessing module: used to pair the active power of the user bus and the active power of each load to form a data set and process the data set;

[0080] A decomposition module: used to initially train three preset base models based on the processed data set, and use the three base models completed in the initial training to decompose the active power of each load to obtain a decomposition result;

[0081] A preliminary aggregation module: used to aggregate the decomposition results using a pre-established perceptron aggregation model and optimize the parameters of the pre-established perceptron aggregation model, and based on the trained perceptron aggregation model, obtain preliminary aggregation data of the active power of each load;

[0082] An output module: used to optimize a pre-established DAE model based on the decomposition results and the preliminary aggregation data. The optimized DAE model is the aggregation model of the non-intrusive load monitoring model, which is used to output the data of the active power of each load.

[0083] In a third aspect, the present invention provides a computing device, including a processor and a storage medium;

[0084] The storage medium is used to store instructions;

[0085] The processor is used to operate according to the instructions to execute the steps of the method described in the first aspect.

[0086] In a fourth aspect, the present invention provides a computer-readable storage medium, on which a computer program is stored, and when the program is executed by a processor, the steps of the method described in the first aspect are implemented.

[0087] Compared with the prior art, the beneficial effects achieved by a non-intrusive load monitoring model aggregation method and system provided by an embodiment of the present invention include:

[0088] The present invention obtains the active power of the user bus and each load; pairs the active power of the user bus and the active power of each load to form a data set, and processes the data set; based on the processed data set, initially trains three preset basic models, and uses the three initially trained basic models to decompose the active power of each load to obtain a decomposition result. Through the initial training of the three basic models, the present invention makes a horizontal comparison to obtain the advantages and disadvantages of each model in load identification for different electrical appliances, and can obtain the optimal model for load identification of each electrical appliance;

[0089] The present invention uses a pre-established perceptron aggregation model to aggregate the decomposition results and optimize the parameters of the pre-established perceptron aggregation model. Based on the trained perceptron aggregation model, preliminary aggregation data of the active power of each load is obtained. Through the trained perceptron aggregation model, the present invention initially fuses the results of different models for load identification of the same electrical appliance. The model with a good identification effect on this electrical appliance will obtain a higher aggregation ratio in the preliminary aggregation to increase the identification accuracy;

[0090] The present invention optimizes a pre-established DAE model based on the decomposition results and the preliminary aggregation data. The optimized DAE model is the aggregation model of the non-intrusive load monitoring model, which is used to output the data of the active power of each load. The present invention can combine three models for identification, has higher universality and generality, can obtain more accurate identification results, aggregates the decomposition results of different loads, and obtains a decomposition result better than that of a single model. Description of the Drawings

[0091] Figure 1 is a flowchart of a method for aggregating a non-intrusive load monitoring model provided in Embodiment 1 of the present invention;

[0092] Figure 2 is the framework of a preset Seq2Point model in a method for aggregating a non-intrusive load monitoring model provided in Embodiment 1 of the present invention;

[0093] Figure 3 is the framework of a preset DAE model in a method for aggregating a non-intrusive load monitoring model provided in Embodiment 1 of the present invention;

[0094] Figure 4 is the framework of a preset Attention and residual network model in a method for aggregating a non-intrusive load monitoring model provided in Embodiment 1 of the present invention. Detailed Embodiment

[0095] The present invention will be further described below with reference to the drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and cannot be used to limit the protection scope of the present invention.

[0096] Example 1:

[0097] As Figure 1 shown, an embodiment of the present invention provides a non-intrusive load monitoring model aggregation method, including:

[0098] Obtain the active power of the user bus and each load;

[0099] Pair the active power of the user bus and the active power of each load to form a data set, and process the data set;

[0100] Based on the processed data set, perform initial training on three preset base models, and use the three base models completed in the initial training to decompose the active power of each load to obtain a decomposition result;

[0101] Use a pre-established perceptron aggregation model to aggregate the decomposition results and optimize the parameters of the pre-established perceptron aggregation model. Based on the trained perceptron aggregation model, obtain the preliminary aggregation data of the active power of each load;

[0102] Based on the decomposition results and the preliminary aggregation data, optimize the pre-established DAE model. The optimized DAE model is the aggregation model of the non-intrusive load monitoring model, which is used to output the data of the active power of each load.

[0103] The specific steps are as follows:

[0104] Step 1: Obtain the active power of the user bus and each load.

[0105] Use the active power monitoring device to extract the active power data of each load and load cluster of the user at a frequency of times / minute, and arrange them in chronological order to obtain the active power of the user bus and the active power of each load.

[0106] Step 2: Pair the active power of the user bus and the active power of each load to form a data set, and process the data set.

[0107] Step 2.1: Pair the active power of the user bus and the active power of each load to form a data set.

[0108] Use a sliding window to obtain the active power data of a specified length, including: set the length of the time series data intercepted by the sliding window to T, and slide and intercept the active power of the obtained user bus and the active power of each load step by step at a step size of 1 in chronological order. The total data length is N,

[0109] Then the data of the active power of the user bus obtained is:

[0110] The active power data of the i-th load is as follows:

[0111] The active power of the user bus and the active powers of each load are paired to form a data set, including:

[0112] Let the data of the active power of the user bus where x n is a vector of T, n = 1, 2, …, N - T + 1,

[0113] The active power data of the i-th load where is a vector of T, n = 1, 2, …, N - T + 1;

[0114] Then is the data set of the pairing of the active power of the user bus and the active powers of each load.

[0115] Step 2.2: Process the data set.

[0116] Perform normalization processing on the data, including: taking out the maximum value of the active power of the user bus, and dividing the active power of the user bus and the active powers of each load by the maximum value of the active power of the user bus respectively.

[0117] Step 3: Conduct initial training on three preset base models based on the processed data set, and decompose the active power of each load using the three base models completed in the initial training to obtain the decomposition results.

[0118] Use the processed data set of the i-th load to train the preset Seq2Point model, and update the model parameters using the reverse optimization algorithm to obtain the trained Seq2Point model;

[0119] Use the processed data set of the i-th load to train the preset DAE model, and update the model parameters using the stochastic gradient descent algorithm to obtain the trained DAE model;

[0120] Use the processed data set of the i-th load to train the preset Attention and residual network model, update the Attention parameters using the BP algorithm to obtain the trained Attention, and based on the trained Attention, update the model parameters using the stochastic gradient descent algorithm to obtain the trained residual network model.

[0121] Specifically, as Figure 2 shown, the preset Seq2Point model includes:

[0122] Input layer: Take The T sampling points are input into a preset Seq2Point model, and the input data tensor is (T, 1);

[0123] Forward GRU layer: Mine the change rules inside the data of the input layer. The hidden layer size is 2T. The input data tensor of the second forward GRU layer is (T, 2T), and the output data tensor is (4T, 1);

[0124] Fully connected layer: Map high-dimensional data to low-dimensional data. The input data tensor is (4T, 1), and the output data tensor is (2T, 1);

[0125] Output layer: Output the normalized power value of the i-th load. The input data tensor is (2T, 1), and the output data tensor is (1, 1).

[0126] Train the preset Seq2Point model, including:

[0127] The activation function used in each unit is the Relu function, expressed as:

[0128]

[0129] The calculation formula of the GRU layer, expressed as:

[0130]

[0131] In formula (2), r t represents the reset gate, W r represents the weight of the reset gate, σ represents the sigmoid function, h t-1 represents the state passed down from the previous neuron, z t represents the state passed down from the previous neuron, W z represents the weight of the update gate, represents the current neuron state, represents the weight value;

[0132] Use the mean square error as the loss function, and the expression is:

[0133]

[0134] In formula (3), MSE represents the loss function, m represents the number of samples of the data for training, represents the active power predicted by the preset Seq2Point model;

[0135] Adopt the reverse optimization algorithm to update the model parameters to make the model converge, and obtain the trained Seq2Point model.

[0136] Specifically, such as Figure 3As shown, the preset DAE model includes:

[0137] Encoder: used to train data and generate a set of feature maps, successively including a convolutional layer, a linear activation function, a max pooling layer, additional convolutional and pooling layers, a fully connected layer, and a ReLU activation function to close the encoder network;

[0138] Decoder: successively including a ReLU activation function, a fully connected layer, additional convolutional and pooling layers, an upsampling layer, a linear activation function, and a convolutional layer.

[0139] Specifically, as Figure 4 shown, the preset Attention and residual network model includes:

[0140] Input layer: used to input the feature x of each sample n ;

[0141] Embedding layer: used to generate a low-dimensional dense vector for feature extraction, and the transformed low-dimensional dense vector is:

[0142] h = WX + b (4)

[0143] In formula (4), h represents the output of the hidden layer of the preset Attention and residual network model, W represents the transformation matrix, b represents the coefficient matrix, and the parameters of W and b are updated based on the BP algorithm during the model training process;

[0144] Encoder, decoder: used to splice the vectors output by the forward GRU and the backward GRU;

[0145] Output layer: used to output the result vector.

[0146] Training the preset Attention includes:

[0147] According to the attention mechanism, using the hidden state sequence [h 0 , h 1 ,..., h n-1 , convert the fixed semantic vector c into a dynamic variable semantic vector c t , and define the parameter vectors Q, K, V, which are expressed as follows:

[0148]

[0149] In formula (5), the parameter vectors Q, K, V are updated based on the BP algorithm during the model training process, then:

[0150] Q = K = V = h = [h 0 , h 1 ,..., h n-1 (6)

[0151] Let the function between the residual connections of the residual network be f(x), then the output of the residual network is g(X) = X + f(X);

[0152] Using cross-entropy as the loss function and the stochastic gradient descent algorithm to update the model parameters, a trained residual network model is obtained.

[0153] Using the three basic models completed in the initial training to perform active power decomposition on the x in the processed dataset n The decomposition results of the three trained basic models for decomposing the i-th load are respectively:

[0154] The decomposition result of the trained Seq2Point model:

[0155] The decomposition result of the trained DAE model:

[0156] The decomposition result of the trained residual network model:

[0157] Through the initial training of the three basic models, this invention can obtain the advantages and disadvantages of each model in load identification for different electrical appliances through horizontal comparison, can obtain the optimal model for each electrical appliance load identification, and can obtain the optimal decomposition result for each electrical appliance load identification.

[0158] Step 5: Use the pre-established perceptron aggregation model to aggregate the decomposition results and optimize the parameters of the pre-established perceptron aggregation model. Based on the trained perceptron aggregation model, preliminary aggregated data of the active power of each load is obtained.

[0159] Form the decomposition results into a decomposition dataset n = 1, 2, …, N - T + 1, where represents the feature, represents the target value;

[0160] Use the pre-established perceptron aggregation model to aggregate the features to obtain:

[0161]

[0162] Use the decomposition dataset to train the pre-established perceptron aggregation model, select the mean squared error as the loss function, and use the stochastic gradient descent algorithm to update the parameters of the pre-established perceptron aggregation model to obtain a trained perceptron aggregation model;

[0163] Based on the trained perceptron aggregation model, preliminary aggregated data of the active power of each load is obtained n = 1, 2, …, N - T + 1.

[0164] In the present invention, the perceptron aggregation model completed through training is used to preliminarily fuse the identification results of different models for the same electrical load. In the preliminary aggregation, the model with good identification effect for the electrical appliance will obtain a higher aggregation ratio to increase the identification accuracy.

[0165] Step 6: Optimize the pre-established DAE model with the decomposition result and the preliminary aggregation data. The optimized DAE model is the aggregation model of the non-intrusive load monitoring model, which is used to output the active power data of each load.

[0166] The model structure of the pre-established DAE model has 7 layers, which are successively:

[0167] The length of the first-layer input layer is determined by the length of the input sequence;

[0168] The second-layer convolutional layer processes the input signal, extracts features to form a feature map, has 8 convolutional kernels with a length of 4, and uses a linear activation function;

[0169] The feature map is expanded through the third layer and then compressed in size through two fully connected layers of the fourth and fifth layers to reduce the dimension of the sequence in order to obtain a compact form of the input sequence, retain the effective information and remove the noise information therein;

[0170] Taking the fifth layer as the midpoint, the sixth and seventh layers are symmetric with the second and fourth layers. The seventh layer is a convolutional layer. The sixth and seventh layers are used to decode the data in the compact form output by the first five layers to obtain an output sequence in the same form as the input sequence.

[0171] Optimizing the pre-established DAE model includes:

[0172] Combining the decomposition result and the preliminary aggregation data to form a training data set;

[0173] Using the training data set to train the pre-established DAE model, selecting the mean square error as the loss function, and adopting the stochastic gradient descent algorithm to update the parameters of the pre-established DAE model to obtain the trained DAE model.

[0174] The present invention can combine three models for identification, has higher universality and generality, can obtain more accurate identification results, aggregate the decomposition results of different loads, and obtain decomposition results better than a single model.

[0175] Embodiment 2:

[0176] The embodiment of the present invention provides a non-intrusive electrical load monitoring multi-model aggregation system, including:

[0177] Acquisition module: used to acquire the active power of the user bus and each load;

[0178] Data preprocessing module: used to pair the active power of the user bus and the active power of each load to form a data set and process the data set;

[0179] Decomposition module: used to initially train three preset base models based on the processed data set, and use the three initially trained base models to decompose the active power of each load to obtain a decomposition result;

[0180] Initial aggregation module: used to aggregate the decomposition results using a pre-established perceptron aggregation model and optimize the parameters of the pre-established perceptron aggregation model, and based on the trained perceptron aggregation model, obtain the initial aggregation data of the active power of each load;

[0181] Output module: used to optimize a pre-established DAE model based on the decomposition result and the initial aggregation data, and the optimized DAE model is the aggregation model of the non-intrusive load monitoring model, which is used to output the data of the active power of each load.

[0182] Embodiment III:

[0183] The embodiment of the present invention provides a computing device, including a processor and a storage medium;

[0184] The storage medium is used to store instructions;

[0185] The processor is used to operate according to the instructions to execute the steps of the method described in Embodiment I.

[0186] Embodiment IV:

[0187] The embodiment of the present invention further provides a computer-readable storage medium, on which a computer program is stored, and when the program is executed by a processor, the steps of the method described in Embodiment I are implemented.

[0188] Those skilled in the art should understand that the embodiments of the present application can be provided as a method, a system, or a computer program product. Therefore, the present application can adopt the form of 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 code.

[0189] This application is described with reference to the flowcharts and / or block diagrams of methods, apparatuses (systems), and computer program products according to embodiments of the present application. It should be understood that each flow and / or block in the flowchart and / or block diagram, and the combination of flows and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to the processors of general-purpose computers, special-purpose computers, embedded processors, or other programmable data processing devices to generate a machine, such that the instructions executed by the processors of the computer or other programmable data processing devices generate means for implementing the functions specified in one or more flows and / or blocks Figure 1 one or more flows and / or blocks Figure 1 or means for implementing the functions specified in one or more blocks.

[0190] These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing device to work in a specific manner, such that the instructions stored in the computer-readable memory generate a manufactured article including instruction means for implementing the functions specified in one or more flows and / or blocks Figure 1 one or more flows and / or blocks Figure 1 or means for implementing the functions specified in one or more blocks.

[0191] These computer program instructions can also be loaded onto a computer or other programmable data processing device, such that a series of operation steps are executed on the computer or other programmable device to generate a computer-implemented process, and thus the instructions executed on the computer or other programmable device provide steps for implementing the functions specified in one or more flows and / or blocks Figure 1 one or more flows and / or blocks Figure 1 or means for implementing the functions specified in one or more blocks.

[0192] The above is only the preferred embodiment of the present invention. It should be noted that for those of ordinary skill in the art, without departing from the technical principle of the present invention, several improvements and modifications can be made, and these improvements and modifications should also be regarded as the protection scope of the present invention.

Claims

1. A non-invasive load monitoring model aggregation method, characterized in that, it includes: Obtain the active power of the user bus and each load; Pair the active power of the user bus and the active power of each load to form a data set, and process the data set; Based on the processed data set, conduct initial training on three preset basic models, and use the three initially trained basic models to decompose the active power of each load to obtain a decomposition result; wherein, the initial training of the three preset basic models includes: Use the data set processed by the i-th load to train the preset Seq2Point model, and update the model parameters using the reverse optimization algorithm to obtain the trained Seq2Point model; Use the data set processed by the i-th load to train the preset DAE model, and update the model parameters using the stochastic gradient descent algorithm to obtain the trained DAE model; Use the data set processed by the i-th load to train the preset Attention and residual network model, update the Attention parameters using the BP algorithm to obtain the trained Attention, and based on the trained Attention, update the model parameters using the stochastic gradient descent algorithm to obtain the trained residual network model; Use the three base models completed in the initial training to perform the active power decomposition of each load on the processed dataset for x n The decomposition results of the three trained base models for the active power of the i-th load are respectively: Decomposition results of the trained Seq2Point model: Decomposition results of the trained DAE model: Decomposition result of the trained residual network model: Use the pre-established perceptron aggregation model to aggregate the decomposition results and optimize the parameters of the pre-established perceptron aggregation model. Based on the trained perceptron aggregation model, obtain the preliminary aggregation data of the active power of each load; Optimize the pre-established DAE model based on the decomposition results and the preliminary aggregation data. The optimized DAE model is the aggregation model of the non-invasive load monitoring model, which is used to output the data of the active power of each load.

2. The non-invasive load monitoring model aggregation method according to claim 1, characterized in that, The obtaining of the active power of the user bus and each load includes: using an active power monitoring device to extract the active power data of each load and load cluster of the user at a frequency of times / minute, arranging them in chronological order to obtain the active power of the user bus and the active power of each load.

3. The non-invasive load monitoring model aggregation method according to claim 1, characterized in that, The composition of the data set includes: Obtain the active power data of a specified length by sliding window, including: setting the length of the time series data intercepted by the sliding window to be T, and sliding and intercepting the obtained active power of the user bus and the active power of each load step by step with a step size of 1 in chronological order, and the total data length is N, Then the data of the user's bus active power obtained is as follows: The data of the active power of the i-th load is as follows: Pair the active power of the user bus and the active power of each load to form a data set, including: The data of the active power of the user bus where x n is a vector T, n = 1, 2, …, N - T + 1 Data of the active power of the i-th load where is a vector of T, and n = 1, 2, …, N - T + 1; Then is a data set that pairs the user bus active power with the active power of each load.

4. The non-invasive load monitoring model aggregation method according to claim 1, characterized in that, The processing of the data set includes: Normalize the data, including: take out the maximum value of the active power of the user bus, and divide the active power of the user bus and the active power of each load by the maximum value of the active power of the user bus respectively.

5. The non-invasive load monitoring model aggregation method according to claim 1, It is characterized in that The obtaining of the preliminary aggregated data of the active power of each load includes: Form the decomposition results into a decomposition data set wherein represents a feature represents a target value Aggregate features using a pre-established perceptron aggregation model Obtain: Using the decomposed data set to train the pre-established perceptron aggregation model, selecting the mean square error as the loss function, and using the stochastic gradient descent algorithm to update the parameters of the pre-established perceptron aggregation model to obtain the trained perceptron aggregation model; Based on the trained perceptron aggregation model, preliminary aggregation data of each load active power is obtained 6. The non-intrusive load monitoring model aggregation method according to claim 5, It is characterized in that The optimization of the pre-established DAE model includes: The decomposition result and the preliminary aggregation data constitute the training data set; Using the training data set to train the pre-established DAE model, selecting the mean square error as the loss function, and using the stochastic gradient descent algorithm to update the parameters of the pre-established DAE model to obtain the trained DAE model.

7. A non-intrusive power load monitoring multi-model aggregation system, It is characterized in that It includes: An acquisition module: used to acquire the active power of the user bus and each load; A data preprocessing module: used to pair the active power of the user bus and the active power of each load to form a data set and process the data set; A decomposition module: used to initially train three preset base models based on the processed data set, and use the three initially trained base models to decompose the active power of each load to obtain a decomposition result; Among them, the initial training of the three preset base models includes: Using the data set processed by the i-th load to train the preset Seq2Point model, and using the reverse optimization algorithm to update the model parameters to obtain the trained Seq2Point model; Using the data set processed by the i-th load to train the preset DAE model, and using the stochastic gradient descent algorithm to update the model parameters to obtain the trained DAE model; Using the data set processed by the i-th load to train the preset Attention and residual network model, using the BP algorithm to update the Attention parameters to obtain the trained Attention, and based on the trained Attention, using the stochastic gradient descent algorithm to update the model parameters to obtain the trained residual network model; Use the three base models completed in the initial training to process the x in the obtained dataset n Perform the active power decomposition of each load. The decomposition results of the three trained base models for the i-th load are respectively: Decomposition results of the trained Seq2Point model: Decomposition results of the trained DAE model: Decomposition result of the trained residual network model: A preliminary aggregation module: used to aggregate the decomposition result using the pre-established perceptron aggregation model and optimize the parameters of the pre-established perceptron aggregation model, and based on the trained perceptron aggregation model, obtain the preliminary aggregated data of the active power of each load; An output module: used to optimize the pre-established DAE model based on the decomposition result and the preliminary aggregated data, and the optimized DAE model is the aggregation model of the non-intrusive load monitoring model, which is used to output the data of the active power of each load.

8. A computing device, It is characterized in that It includes a processor and a storage medium; The storage medium is used to store instructions; The processor is used to operate according to the instructions to execute the steps of the method according to any one of claims 1 to 6.

9. A computer-readable storage medium, on which a computer program is stored, It is characterized in that When the program is executed by the processor, it implements the steps of the method according to any one of claims 1 to 6.

Images