A deep learning-based non-intrusive load decomposition method and system

Abstract

The application provides a non-intrusive load decomposition method and system based on deep learning, which comprises the following steps: preprocessing historical load operation power data of a user side to obtain aggregated power data; and establishing a deep learning model, including a sequence-to-sequence model and a sequence-to-point model; wherein the sequence-to-sequence model extracts features of a power sequence of the preprocessed aggregated power data as global features; and the global features output by the sequence-to-sequence model are divided into time windows by adopting a sliding window method according to a time dimension; and the sequence-to-point model outputs a median value of a target sequence corresponding to each time window to predict the target sequence point by point.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of non-intrusive load decomposition, and particularly relates to a non-intrusive load decomposition method and system based on deep learning. BACKGROUND

[0002] The statements in this section merely provide background information related to the present application and do not necessarily constitute the prior art.

[0003] The energy demand of buildings accounts for more than 40% of the total energy consumption in the world, and at the same time, with the development of urbanization, the demand for energy of buildings will further increase. Therefore, the energy management on the demand side is of great significance to achieve the goal of energy saving and emission reduction.

[0004] The non-intrusive load monitoring technology receives the total load operation data in the lower circuit by installing sensors at the power inlet of the user, and then uses data mining technology to decompose the power consumption of the target electrical appliance. This technology provides detailed and accurate data support for efficient monitoring and scheduling of electric energy, and has far-reaching application value for demand-side energy management.

[0005] At present, the load power decomposition algorithm based on deep learning mainly includes sequence-to-sequence algorithm and sequence-to-point algorithm.

[0006] Among them, the sequence-to-sequence algorithm is similar to a language translation model, taking the total load consumption sequence as input and outputting the power sequence of the target electrical appliance. Although this model can effectively separate the load, it cannot accurately identify the state edge of the load.

[0007] The sequence-to-point algorithm converts the output sequence of the model into the midpoint of the target sequence, effectively improving this problem and reducing the decomposition error. However, since the entire model needs to be run once when calculating each point in the target sequence, this algorithm greatly increases the demand for computing resources.

[0008] Therefore, the sequence-to-sequence model based on deep learning currently has the problems of unclear load switch state recognition and large amount of calculation of the sequence-to-point model. SUMMARY

[0009] In order to overcome the shortcomings of the prior art, the present application provides a non-intrusive load decomposition method based on deep learning, which combines sequence-to-point and sequence-to-sequence load decomposition algorithms. In order to achieve the above purpose, one or more embodiments of the present application provide the following technical solutions:

[0010] In a first aspect, a deep learning-based non-intrusive load decomposition method is disclosed, comprising:

[0011] Preprocessing user-side historical load operation power data to obtain aggregated power data;

[0012] Establishing a deep learning model, including a sequence-to-sequence model and a sequence-to-point model;

[0013] The sequence-to-sequence model extracts features of power sequences from the preprocessed aggregated power data as global features;

[0014] The global features output by the sequence-to-sequence model are divided into time windows using a sliding window method according to the time dimension;

[0015] The sequence-to-point model outputs the median of the target sequence corresponding to each time window to predict the target sequence point by point.

[0016] As a further technical solution, the preprocessing of the user-side historical load operation data is specifically:

[0017] First, align the total historical load operation power data and the historical operation power data of each load according to the time stamp;

[0018] Then, backward data filling is performed on the historical data of a single load according to the time stamp of the total power to ensure data integrity;

[0019] Then, for the missing data in the filled data, the time point of the missing data is removed;

[0020] Finally, the aggregated power data and the power data of a single load after the removal processing are standardized.

[0021] As a further technical solution, after preprocessing the user-side historical load operation data, the preprocessed data is divided into a training set and a test set according to different rooms for training and selection of the deep learning model.

[0022] As a further technical solution, the preprocessed aggregated power data is input into the sequence-to-sequence model, and the data is iteratively input into the sequence-to-sequence model using a sliding window method.

[0023] As a further technical solution, the sequence-to-sequence model includes a time convolution-based deep neural network, which processes the iteratively input data and outputs window features of a certain time length and dimension.

[0024] As a further technical solution, the obtained window features are iteratively input to the sequence-to-point model again in a sliding window manner, and the final output is a normalized power prediction value of the target disaggregated appliance.

[0025] As a further technical solution, the normalized power prediction value is de-normalized to obtain a real target appliance power prediction value.

[0026] As a further technical solution, the sequence-to-sequence model includes 7 time convolution layers, of which the 2nd to 4th are empty convolution layers, and the parameters in each convolution represent the output channel size, the convolution kernel size, the step, and the empty convolution parameter in sequence.

[0027] Each convolution layer is followed by a ReLU activation function, and a window feature of a certain time length and dimension is output.

[0028] The sequence-to-point model includes three convolution layers with ReLU activation functions, two linear layers, and two Dropout layers.

[0029] In a second aspect, a non-intrusive load disaggregation system based on deep learning is disclosed, comprising:

[0030] A preprocessing module configured to preprocess user-side historical load operating power data to obtain aggregated power data;

[0031] A deep learning model establishment module configured to establish a deep learning model, including a sequence-to-sequence model and a sequence-to-point model;

[0032] The sequence-to-sequence model extracts power sequence features from the preprocessed aggregated power data as global features.

[0033] The global features output by the sequence-to-sequence model are divided into time windows using a sliding window method according to the time dimension.

[0034] The sequence-to-point model outputs the median of the target sequence corresponding to each time window to predict the target sequence point by point.

[0035] As a further technical solution, two sliding window modules are further included, namely a first sliding window module and a second sliding window module.

[0036] The first sliding window module processes the iteratively input data, and then inputs it to the sequence-to-sequence model.

[0037] The second sliding window module processes the data output by the sequence-to-sequence model, and then inputs it to the sequence-to-sequence model.

[0038] The one or more technical solutions have the following beneficial effects:

[0039] The framework of the algorithm is based on a full convolution model, and the framework of the model is divided into two stages. In the first stage, a sequence-to-sequence model architecture is used to extract the features of the input power sequence, and in the second stage, a sequence-to-point model architecture is used. First, according to the features output in the previous stage, a sliding window method is used to divide smaller time windows in the time dimension, then each time window is input into the second stage model, and finally the median of the target sequence corresponding to each small time window is output. In the first stage, the global features are extracted, which potentially expands the size of the input window in the stage, and in the second stage, the sequence-to-point algorithm is used to predict the target sequence point by point. The algorithm effectively combines the advantages of the sequence-to-sequence algorithm and the sequence-to-point algorithm, that is, it retains the efficiency of the former and the accuracy of the prediction of the latter.

[0040] Advantages of the additional aspects of the application will be partially given in the following description, partially will become apparent from the following description, or will be understood by the practice of the application. BRIEF DESCRIPTION OF DRAWINGS

[0041] The accompanying drawings, which form a part of this specification, are included to provide a further understanding of the application, and are incorporated by reference herein. The embodiments depicted in the drawings are provided by way of explanation of the application, and are not meant to limit the application.

[0042] Figure 1 Data processing flowchart used for the embodiments of the application;

[0043] Figure 2 Algorithm flowchart used for the embodiments of the application;

[0044] Figure 3 Sequence-to-sequence model instance used for the embodiments of the application;

[0045] Figure 4 Sequence-to-point model instance used for the embodiments of the application. DETAILED DESCRIPTION

[0046] It should be noted that the following detailed description is exemplary in nature and is intended to provide further description of the application. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which the application belongs.

[0047] It should be noted that the terms used herein are only for the purpose of describing specific embodiments, and are not intended to limit the exemplary embodiments according to the application.

[0048] In the case of no conflict, the embodiments in the application and the features in the embodiments can be combined with each other.

[0049] Embodiment one

[0050] The embodiment discloses a deep learning-based non-invasive load decomposition method, comprising:

[0051] First, the historical load operation data in the family is processed, then the proposed deep learning model is established, and the processed data is optimized and trained to the established model, and finally the optimized best model is used for the actual load decomposition task.

[0052] In the embodiment, the data processing process is as shown in Figure 1 The total load operation power data and the historical operation power data of each load are aligned according to the time stamp. Since the single load has sparsity in the data acquisition process, the historical data of the single load is backward data filled according to the time stamp of the total power to ensure the integrity of the data. Then, for the missing data in the filled data, the time point of the missing data is removed. Then, the aggregated power data and the power data of the single load are standardized according to formula (1), specifically:

[0053]

[0054] Where x mean represents the mean of the data, and x std represents the standard deviation of the data.

[0055] The above standardized data is two groups, that is, the aggregated power sequence and the power sequence of the single appliance. The aggregated power sequence is the input of the model, and the power of the single load is the target of the model prediction. The purpose of the algorithm is to input all the power sequences of the appliances to the model, and then the model outputs the power sequence of the single appliance we want.

[0056] Finally, the obtained data is divided into a training set and a test set according to different rooms, for model training and selection.

[0057] The algorithm implementation framework of the application is as shown in Figure 2 The framework mainly includes two deep learning models, two sliding window modules and a reverse standardization process. The above deep learning model includes a sequence-to-sequence model and a sequence-to-point model, and the two sliding window modules are a first sliding window module and a second sliding window module.

[0058] The processed aggregated power data is used as the input of the sequence-to-sequence model, the first sliding window module iteratively inputs the data to the sequence-to-sequence model in a sliding window manner, and the window size is w1 and the step size is s1 each time.

[0059] The above-mentioned sliding window method iteratively inputs data to the sequence-to-sequence model is a general data input method when load decomposition. The data is divided into several window size data, so that the length of the input data is the same, let the sequence-to-sequence model does not need to input a large amount of data at a time. Because the power data is usually aggregated in time dimension, it cannot be input to the sequence-to-sequence model at a time.

[0060] The first stage of the sequence-to-sequence model is a sequence-to-sequence time convolution-based deep neural network, and the data is output through this stage. The window feature has a time length of w1 and a dimension of d. The sequence-to-sequence model instantiation is as shown in Figure 3 .

[0061] The neural network is composed of several convolutional layers as shown in Figure 3 . After inputting a sequence, the convolutional layer will perform convolution operation on the sequence. After all the convolution operations, a feature vector will be output, which is the feature of the input window, i.e. the window feature. The neural network can be regarded as a feature extractor, which inputs a window length data and outputs the feature of the window length data to obtain the window feature.

[0062] Then the second sliding window module iteratively inputs the obtained window feature to the second stage sequence-to-point model using the sliding window method. The window size used here is w2 and the step size is s2. In this stage, because it is a sequence-to-point model, only sequence-to-point input method can be used. The model instantiation is as shown in Figure 4 . The output of the final model is the normalized power prediction value of the target decomposition appliance.

[0063] The input of the sequence-to-point model is the window feature output by the previous stage sequence-to-sequence model. The sequence-to-point model includes convolutional layers, linear layers and Dropout layers as shown in Figure 4 . The convolutional layer further extracts features from the window feature output by the previous stage. Then the linear layer is used for feature transformation to output the power of the target appliance. Since the power of the target appliance has been normalized, the loss is calculated based on the normalized data when calculating the model output. Therefore, the output here is the normalized power of the target appliance, i.e. the prediction value output by the model is already normalized.

[0064] Therefore, the prediction value needs to be de-normalized to obtain the real target appliance power prediction value. The de-normalization process is as shown in formula (2):

[0065]

[0066] where represents the mean of the target appliance power data in the training set, a standard deviation of the target appliance power data in the training set.

[0067] In a more detailed embodiment, the implementation details of the scheme are as follows: in the case where the window size w1 of the first sliding window module is set to 1199 and the window size w2 of the second sliding window module is set to 599, s1 should be 601 and s2 should be 1.

[0068] The sequence-to-sequence model of the first stage includes 7 time convolution layers, of which the 2nd to 4th are empty convolution layers. The parameters in each convolution represent the output channel size, the convolution kernel size, the step, and the empty convolution parameter in sequence. Here, the role of each convolution layer is to perform feature extraction (convolution operation), and the input of each layer is the output of the previous layer, and the obtained is a feature vector. Each convolution layer is followed by a ReLU activation function, and the specific parameter settings are as shown in Table 1. Figure 3 After the first stage, a window feature with a time length of 1199 and a dimension of 64 is output.

[0069] The model of the second stage includes three convolution layers with ReLU activation functions, two linear layers, and two Dropout layers. The specific parameter settings are as shown in Table 2. Figure 4 The above convolution layers perform convolution operations on the input, the linear layers perform linear transformations, the activation functions perform nonlinear transformations, and the output is a feature vector.

[0070] In the training phase of the model, the initialization parameters of the model need to be set. In this scheme, all weight parameters are set using the Kaiming initialization. In addition, the learning rate (Learning Rate), batch size (Batch Size), and Dropout parameters used in the training process are set, and the Adam optimization algorithm is used to update the model parameters. This scheme uses the Mean Square Error (MSE) as shown in equation (3) as the loss function of the model, and uses the Mean Absolute Error (MAE) as shown in equation (4) to evaluate the performance of the model.

[0071]

[0072]

[0073] where N represents the time length of the sample, y i represents the predicted value of the model, represents the actual consumption value of the target appliance.

[0074] Example Two

[0075] The embodiment aims to provide a computer device, comprising a memory, a processor and a computer program stored on the memory and executable on the processor, wherein the processor implements the steps of the above method when executing the program.

[0076] Embodiment three

[0077] The embodiment aims to provide a computer readable storage medium.

[0078] A computer readable storage medium, having a computer program stored thereon, wherein the program is executable by a processor to implement the steps of the above method.

[0079] Embodiment four

[0080] The embodiment aims to provide a deep learning-based non-invasive load decomposition system, comprising:

[0081] A preprocessing module configured to preprocess user-side historical load operation power data to obtain aggregated power data;

[0082] A deep learning model establishing module configured to establish a deep learning model, including a sequence-to-sequence model and a sequence-to-point model;

[0083] The sequence-to-sequence model extracts features of power sequences of the preprocessed aggregated power data as global features;

[0084] The global features output by the sequence-to-sequence model are divided into time windows by using a sliding window method in the time dimension;

[0085] The sequence-to-point model outputs the median of the target sequence corresponding to each time window to predict the target sequence point by point.

[0086] As a further technical solution, it further comprises two sliding window modules, namely a first sliding window module and a second sliding window module;

[0087] The first sliding window module processes the iteratively input data, and then inputs the data to the sequence-to-sequence model;

[0088] The second sliding window module processes the data output by the sequence-to-sequence model, and then inputs the data to the sequence-to-sequence model.

[0089] The steps involved in the apparatuses of the above embodiments two, three and four correspond to the method of embodiment one, and the specific implementation can refer to the relevant description of embodiment one. The term "computer readable storage medium" should be understood as including a single medium or multiple media of one or more instruction sets; it should also be understood as including any medium capable of storing, encoding or carrying the instruction set for execution by the processor and causing the processor to perform any of the methods in the present application.

[0090] Those skilled in the art should understand that each module or step of the present application described above can be realized by a general computer device, alternatively, they can be realized by program codes executable by a computing device, so that they can be stored in a storage device for execution by a computing device, or they can be respectively made into each integrated circuit module, or a plurality of modules or steps among them can be made into a single integrated circuit module to realize. The present application is not limited to any specific combination of hardware and software.

[0091] Although the specific embodiments of the present application are described above in combination with the drawings, it is not a limitation on the scope of protection of the present application, and those skilled in the art should understand that various modifications or changes made by those skilled in the art on the basis of the technical solutions of the present application without creative labor are still within the scope of protection of the present application.

Claims

1. A deep learning-based non-intrusive load decomposition method, characterized by, The method comprises the following steps: Preprocessing user-side historical load operation power data to obtain aggregated power data; Establishing a deep learning model, including a sequence-to-sequence model and a sequence-to-point model; The sequence-to-sequence model extracts features of power sequences from the preprocessed aggregated power data as global features; The preprocessed aggregated power data is input into the sequence-to-sequence model, and the data is iteratively input into the sequence-to-sequence model in a sliding window manner; The global features output by the sequence-to-sequence model are divided into time windows in a time dimension by using a sliding window method; The sequence-to-sequence model includes a sequence-to-sequence time convolution-based deep neural network, which processes the iteratively input data and outputs window features of a certain time length and dimension; The sequence-to-point model outputs the median of the target sequence corresponding to each time window to predict the target sequence point by point; The obtained window features are iteratively input into the sequence-to-point model in a sliding window manner, and the final output is the normalized power prediction value of the target decomposed appliance.

2. The deep learning based non-intrusive load decomposition method of claim 1, wherein, The preprocessing of the user-side historical load operation data specifically comprises the following steps: Firstly, align the total historical load operation power data and the historical operation power data of each load according to the time stamp; Then, backward data filling is performed on the historical data of a single load according to the time stamp of the total power to ensure the integrity of the data; Then, the time points of the missing data are removed for the data missing in the filled data; Finally, the aggregated power data and the power data of a single load after the removal processing are standardized.

3. The deep learning based non-intrusive load decomposition method of claim 1, wherein, After preprocessing the user-side historical load operation data, the preprocessed data is divided into a training set and a test set for training and selection of the deep learning model.

4. The deep learning based non-intrusive load decomposition method according to any one of claims 1-3, characterized in that, Further comprising: The normalized power prediction value is de-normalized to obtain the real target appliance power prediction value.

5. The deep learning based non-intrusive load decomposition method according to any one of claims 1-3, characterized in that, The sequence-to-sequence model includes 7 time convolution layers, of which the 2nd to 4th are empty convolution layers, and the parameters in each convolution represent the output channel size, convolution kernel size, step, and empty convolution parameter in sequence; Each convolution layer is followed by a ReLU activation function, which outputs window features of a certain time length and dimension; The sequence-to-point model includes three convolution layers with ReLU activation functions, two linear layers, and two Dropout layers.

6. A deep learning based non-intrusive load disaggregation system, implementing the deep learning based non-intrusive load disaggregation method of any one of claims 1-5, characterized in that, The method comprises the following steps: A preprocessing module is configured to preprocess user-side historical load operation power data to obtain aggregated power data; A deep learning model establishment module is configured to establish a deep learning model, including a sequence-to-sequence model and a sequence-to-point model; The sequence-to-sequence model extracts features of power sequences from the preprocessed aggregated power data as global features; The global features output by the sequence-to-sequence model are divided into time windows in a time dimension by using a sliding window method; The sequence-to-point model outputs the median of the target sequence corresponding to each time window to predict the target sequence point by point.

7. A computer apparatus comprising a memory, a processor, and a computer program stored on the memory and loadable on the processor, characterized in that, The processor executes the program to realize the steps of the method of any one of claims 1-5.

8. A computer-readable storage medium having stored thereon a computer program, characterized in that, The program, when executed by the processor, performs the steps of the method of any of claims 1-5.

Images