Non-intrusive load decomposition method based on 1D convolutional neural network and LSTM

Abstract

The present invention discloses a non-intrusive load decomposition method based on a combination of 1D convolutional neural network and LSTM, including: S1 collecting active power time series data of total load and various single loads within a period of time as sample data; S2 training corresponding neural network models for various single loads respectively, specifically: taking the active power time series data of the total load as input and the active power time series data of various single loads as output, training the neural network model combining 1D convolutional neural network and LSTM to obtain the neural network model corresponding to various single loads; S3 inputting the active power data of the total load into the trained neural network model, then decomposing the active power data of the corresponding single load. The present invention does not need to use an event monitoring mechanism to detect the change point of the load curve, which greatly simplifies the load decomposition process.

Description

Technical Field

[0001] The present invention belongs to the technical field of power load decomposition, and specifically relates to a non-intrusive load decomposition method based on the combination of 1D convolutional neural network and LSTM. Background Art

[0002] In recent years, with the construction of ubiquitous power Internet of Things and the advent of the era of smart electricity consumption, smart meters are expected to become mainstream products in various terminals of the State Grid. By analyzing the electricity consumption data collected in real time by smart meters, the electricity consumption patterns and power consumption of loads can be accurately understood in real time, so as to guide users to use electricity more reasonably and achieve the goal of demand-side response. In addition, the use of power resources can be monitored in real time through smart meters, and measures can be taken to reduce energy waste when necessary, which has great research significance for the rational use of power resources.

[0003] Non-intrusive load decomposition is to decompose the total power data into the power consumption of a single appliance without installing a large number of sensors and monitoring equipment. Non-intrusive load decomposition only requires adding a non-intrusive load decomposition module to the user's total electricity meter, and decomposing the operating status and related parameter information of different loads of the user based on the time series values ​​obtained by measurement, so as to realize online monitoring of load power consumption. This not only saves a large number of monitoring equipment and cumbersome sensor devices, greatly reduces economic costs, and effectively avoids the problems of poor practical operability, high implementation and maintenance costs, and inconvenient equipment maintenance and repair caused by intrusive load monitoring. At the same time, it also improves the stability and reliability of system monitoring, and will not cause too much disturbance to the production and life of users.

[0004] Non-intrusive load decomposition is mainly aimed at power grid companies. This technology is more inclined to understand the power consumption of various electrical appliances, and is not concerned with the exact moment when the appliances are turned on / off. At present, the conventional method is to combine the transient and steady-state characteristics of the power system and use pattern recognition methods to decompose non-intrusive power loads. However, these methods generally have low decomposition accuracy for power loads with similar characteristics. Load decomposition technology is still a very challenging problem. Research in this field is still in an immature stage, and there are still major technical difficulties to be overcome. Summary of the invention

[0005] The purpose of the present invention is to provide a non-invasive load decomposition method based on a combination of a 1D convolutional neural network and LSTM. The method of the present invention can decompose the total power information obtained by the sensors installed at the load inlet to obtain the active power information of each power load at different times.

[0006] The non-intrusive load decomposition method based on the combination of 1D convolutional neural network and LSTM provided by the present invention includes:

[0007] S1 collects the total load and active power time series data of various single loads within a period of time as sample data;

[0008] S2 trains the corresponding neural network models for each type of single load. Specifically, the active power time series data of the total load is used as input, and the active power time series data of each type of single load is used as output. The neural network model combining 1D convolutional neural network and LSTM is trained to obtain the neural network models corresponding to each type of single load.

[0009] The neural network model combining the 1D convolutional neural network and the LSTM includes an input layer, a hidden layer and an output layer;

[0010] The input layer is used to receive a one-dimensional vector, i.e., active power time series data of the total load;

[0011] The hidden layer includes 3 convolutional layers, 2 LSTM layers and 2 fully connected layers; wherein, the first convolutional layer is composed of 32 convolutional kernels with a dimension of 5, and its convolution step is 1, and the BatchNormalization algorithm is used to normalize the input of the activation function; the second convolutional layer is composed of 32 convolutional kernels with a dimension of 5, and its convolution step is 1, and the BatchNormalization algorithm is used to normalize the input of the activation function; the third convolutional layer is composed of 16 convolutional kernels with a dimension of 4, and its convolution step is 1, and the BatchNormalization algorithm is used to normalize the input of the activation function; the first LSTM layer uses a size of 64 dimensions for output; the second LSTM layer uses a size of 128 dimensions for output; the second LSTM layer is followed by a fully connected layer with a dimension of 64, which fully connects the result obtained in the previous layer, and then calls the ReLU activation function, and then performs a full connection with a dimension of 1. After a linear change, the final output, i.e., the active power information of a specific load, is obtained;

[0012] The output layer outputs the active power information of the load corresponding to the neural network;

[0013] S3 inputs the active power data of the total load into the trained neural network model, and decomposes the active power data of the corresponding single load.

[0014] Furthermore, the activation function in the first fully connected layer adopts the ReLU activation function, and the activation functions in the convolutional layer, LSTM layer and the second fully connected layer are all linear functions.

[0015] Preferably, there is a Dropout layer with a parameter of 0.2 between every two convolutional layers in the hidden layer;

[0016] There is also a Dropout layer with a parameter of 0.2 between the first LSTM layer and the second LSTM layer;

[0017] There is a Dropout layer with a parameter of 0.5 between the second LSTM layer and the fully connected layer, and between the two fully connected layers.

[0018] The present invention has the following advantages and beneficial effects:

[0019] The present invention does not need to use an event monitoring mechanism to detect the change point of the load curve, which greatly simplifies the load decomposition process. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 It is a schematic diagram of the process of the present invention;

[0021] Figure 2 It is a schematic diagram of the network model adopted in the specific implementation method. DETAILED DESCRIPTION

[0022] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention will be further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0023] Before describing the specific implementation method, it needs to be explained that the load mentioned below and in the whole text refers to the electrical equipment.

[0024] Active power can be used as one of the main features to distinguish different loads. The present invention uses the active power of the total load on the user side and the active power of each single load as the input and output of the neural network model, that is, the input layer is the active power data of the total load on the user side, and the output layer is the decomposed single load active power data. The active power of the total load and the active power of each single load are used to train the constructed neural network model combining 1D convolutional neural network and LSTM. After the training is completed, it is only necessary to input the total load active power data into the trained neural network model to output the power consumption information of a specific single load.

[0025] Assume that the household power consumption data for each single load in period t is Here, it is assumed that a total of m loads are used as the load types in the monitoring environment, and the power consumption data of each load is accumulated as the total power consumption P t , that is, the active power data of the total load. The time series of the active power of the total household load (P1, P2, ..., P t ),(P t+1 , P t+2 ,…,P 2t ),…,(P (n-1)×t+1, P (n -1)×t+2,…,P n×t ), as the input sample of the constructed neural network model, the subscripts of each active power in the above active power time series represent the time period, that is, P i Represents the total load active power corresponding to time period i, i = 1, 2, ... n × t; the active power of a single load that needs to be decomposed within the corresponding time interval n × t As the output sample of the neural network model. The constructed neural network model is continuously trained using sample data until the accuracy requirements are met. For each type of power load, the corresponding neural network model is trained in the same way. Each trained network model is equivalent to the feature extractor of the specific load. For any total load power curve, as long as the total load power curve is input, the network model can decompose the power consumption information of the specific load from the total power curve.

[0026] The goal achieved by the present invention is to use the total load active power data within a certain period of time as input data, and output the active power curve of a specific load through a neural network that combines a trained 1D convolutional neural network and LSTM. In order to perform load decomposition, the designed neural network model includes an input layer, a hidden layer, and an output layer. The input layer uses the active power of the total load as the input of the neural network. The output layer has only one node, which corresponds to a certain load. Given the electricity consumption data within a certain period of time, the output result is the electricity consumption information of the specific load within the period of time. The present invention needs to be trained separately for each load. A neural network model only corresponds to one household load. As many neural network models of this structure as there are loads need to be trained, and the output of the neuron is the electricity consumption information of the load.

[0027] The following will provide a method for constructing a neural network model that combines 1D convolutional neural network and LSTM.

[0028] The problem that needs to be solved for non-intrusive load decomposition is to identify the components of the load from the collected total load power consumption curve. The present invention uses an improved neural network to solve this problem.

[0029] like Figure 1 As shown in the figure, the constructed neural network model structure is a combination of 1D convolutional neural network and LSTM. The neural network model is used to train the active power of the load and extract the corresponding active power curve features. The neural network model structure of each type of load is further constructed as an identifier for extracting the corresponding type of load. After the training is completed, the model is applied to the reasoning scenario to decompose the input total load power consumption curve, so as to achieve the purpose of decomposing a certain load from each signal.

[0030] Specifically, given a time window x(t) with a length of t starting from a certain time point, the time window x(t) contains the total active power information of K loads, using model g (j) (x(t),θ) estimates the power curve of the jth load under the corresponding time series, where θ is the network model parameter, which is automatically determined by model training:

[0031]

[0032] In formula (1), Represents the activation signal subnetwork from the input layer to the output layer of the neural network model, It means the output signal obtained after the active power of the jth load in the time window x(t) passes through the neural network. The output layer is activated by the activation function σ, and the activated output is mapped to the valid data range. The activation function σ uses the ReLU function, that is, σ(γ)=max(0,γ), and γ represents the input of the activation function. Substituting formula (1) into it, we have:

[0033]

[0034] ω (j) (t) represents the output of the activation function.

[0035] In the training phase, (x(t), ω(t)) is called a training segment with K data samples. A non-overlapping window is used to extract the training segment from the mixed signal time series (x(t), ω(t)) so that the input of the training set is:

[0036] x=((x1,x2,...,x k ),(x k+1 ,x k+2 ,...,x 2k ),...,(x (n-1)*k+1 ,x (n-1)*k+2 ,...,x nk ))(3)

[0037] In formula (3), (x1,x2,...,x k ),(x k+1 ,x k+2 ,...,x 2k ),...,(x (n-1)*k+1 ,x (n-1)*k+2 ,...,x nk ) is the time series of the total active power of household load mentioned above (P1, P2, ..., P t ),(P t+1 , P t+2 ,…,P 2t ),…,(P(n-1)×t+1 , P (n-1)×t+2 ,…,P n×t ).

[0038] The corresponding single load output power Ω is recorded as:

[0039] Ω=(ω(k),ω(2k),...,ω(n*k)) (4)

[0040] Then the output error relative to the true active power of a single load is calculated, and the mean square error (MSE) is used as the loss function to train the network model.

[0041] The structure of the neural network model constructed by the present invention based on the combination of 1D convolutional neural network and LSTM is described as follows.

[0042] See Figure 2 As shown in FIG. 1 , the neural network includes an input layer, a hidden layer and an output layer. The input layer is the total active power time series of the load within a specified timestamp, which is a one-dimensional vector containing the total active power information sampled at a certain frequency.

[0043] The hidden layer contains 3 convolutional layers, 2 LSTM layers and 2 fully connected layers. The first convolutional layer consists of 32 convolution kernels with a dimension of 5 and a convolution step of 1. The Batch Normalization algorithm (i.e. Figure 1 The BN in the activation function is used to normalize the input of the activation function to solve the impact of input data offset and increase. The activation function uses a linear function. The second convolution layer is also composed of 32 convolution kernels with a dimension of 5, and its convolution step is 1. The BatchNormalization algorithm is also used to normalize the input of the activation function, and the activation function uses a linear function. The third convolution layer is composed of 16 convolution kernels with a dimension of 4, and its convolution step is 1. The Batch Normalization algorithm is used to normalize the input of the activation function, and the activation function uses a linear function. The first LSTM layer uses a size of 64 dimensions for output, and the activation function uses a linear function. The second LSTM layer uses a size of 128 dimensions for output, and the activation function uses a linear function. After the second LSTM layer, there is a fully connected layer with a dimension of 64. The result obtained from the previous LSTM layer is fully connected, and then the ReLU activation function is called. Then a full connection with a dimension of 1 is performed. After a linear change, the final output is obtained, that is, the active power information of a specific load.

[0044] The output layer outputs the active power information of the load corresponding to the neural network.

[0045] There is a Dropout layer with a parameter of 0.2 between every two convolutional layers in the hidden layer. There is also a Dropout layer with a parameter of 0.2 between the first LSTM layer and the second LSTM layer. There is a Dropout layer with a parameter of 0.5 between the second LSTM layer and the fully connected layer, and between the two fully connected layers. The function of the Dropout layer is to remove the training units of the neural network from the network according to a certain probability during the deep learning training process. Since it is randomly discarded, a different network is trained each time batch_size samples are randomly selected.

[0046] The deep learning model has a powerful learning and expression ability. Although the data collected from the smart meter is the composite data after the superposition of various loads, and the relationship between the various decomposed data is very complex and there is no objective law to follow, the deep learning model uses multiple nonlinear hidden layers to continuously learn the deep features of the object and re-express the input data. Not only that, these abstract features also express the relationship between input and output very well. Compared with traditional methods, when using deep learning methods to process time series data, there is no need to distinguish between transient and steady-state processes of the load, and the end-to-end load decomposition process can be achieved without detecting the switching events of the load.

[0047] Those of ordinary skill in the art will appreciate that, in the various method embodiments of the present invention, the sequence numbers of the steps described cannot be used to limit the order of the steps. For those of ordinary skill in the art, without paying creative labor, the order of the steps is also within the protection scope of the present invention. The embodiments described here are to help readers understand the implementation method of the present invention, and it should be understood that the protection scope of the present invention is not limited to such special statements and embodiments. Those of ordinary skill in the art can make various other specific variations and combinations that do not deviate from the essence of the present invention based on the technical enlightenment disclosed by the present invention, and these variations and combinations are still within the protection scope of the present invention.

Claims

1. A non-intrusive load decomposition method based on a combination of 1D convolutional neural network and LSTM, characterized by: include: S1 collects the total load and active power time series data of various single loads within a period of time as sample data; S2 trains the corresponding neural network models for each type of single load. Specifically, the active power time series data of the total load is used as input, and the active power time series data of each type of single load is used as output. The neural network model combining 1D convolutional neural network and LSTM is trained to obtain the neural network models corresponding to each type of single load. The neural network model combining the 1D convolutional neural network and the LSTM includes an input layer, a hidden layer and an output layer; The input layer is used to receive a one-dimensional vector, i.e., active power time series data of the total load; The hidden layer includes 3 convolutional layers, 2 LSTM layers and 2 fully connected layers; wherein, the first convolutional layer is composed of 32 convolutional kernels with a dimension of 5, and its convolution step is 1, and the BatchNormalization algorithm is used to normalize the input of the activation function; the second convolutional layer is composed of 32 convolutional kernels with a dimension of 5, and its convolution step is 1, and the BatchNormalization algorithm is used to normalize the input of the activation function; the third convolutional layer is composed of 16 convolutional kernels with a dimension of 4, and its convolution step is 1, and the BatchNormalization algorithm is used to normalize the input of the activation function; the first LSTM layer uses a size of 64 dimensions for output; the second LSTM layer uses a size of 128 dimensions for output; the second LSTM layer is followed by a fully connected layer with a dimension of 64, which fully connects the result obtained in the previous layer, and then calls the ReLU activation function, and then performs a full connection with a dimension of 1. After a linear change, the final output, i.e., the active power information of a specific load, is obtained; The output layer outputs the active power information of the load corresponding to the neural network; S3 inputs the active power data of the total load into the trained neural network model, and decomposes the active power data of the corresponding single load; The activation function in the first fully connected layer uses the ReLU activation function, and the activation functions in the convolutional layer, LSTM layer, and the second fully connected layer are all linear functions; There is a Dropout layer with a parameter of 0.2 between every two convolutional layers in the hidden layer; There is also a Dropout layer with a parameter of 0.2 between the first LSTM layer and the second LSTM layer; There is a Dropout layer with a parameter of 0.5 between the second LSTM layer and the fully connected layer, and between the two fully connected layers; Given a time window x(t) with a length of t starting from a certain time point, the time window x(t) contains the total active power information of K loads, using model g (j) (x(t),θ) estimates the power curve of the jth load under the corresponding time series, where θ is the network model parameter, which is automatically determined by model training: In formula (1), Represents the activation signal subnetwork from the input layer to the output layer of the neural network model, It means the output signal obtained after the active power of the jth load in the time window x(t) passes through the neural network. The output layer is activated by the activation function σ, and the activated output is mapped to the valid data range. The activation function σ uses the ReLU function, that is, σ(γ) = max(0,γ), where γ represents the input of the activation function. Substituting formula (1) into it, we have: ω (j) (t) represents the output of the activation function; In the training phase, (x(t), ω(t)) is called a training segment with K data samples; a non-overlapping window is used to extract the training segment from the mixed signal time series (x(t), ω(t)), so that the input of the training set is: x=((x1,x2,…,x k ),(x k+1 , x k+2 ,…,x 2k ),…,(x (n-1)*k+1 , x (n −1)*k+2,…,x nk ))(3) In formula (3), (x1,x2,...,x k ),(x k+1 ,x k+2 ,...,x 2k ),...,(x (n-1)*k+1 ,x (n -1)*k+2,...,x nk ) is the time series of the total active power of household load (P1, P2, ..., P t ),(P t+1 , P t+2 ,…,P 2t ),…,(P (n-1)×t+1 , P (n -1)×t+2,…,P n×t ); The corresponding single load output power Ω is recorded as: Ω=(ω(k),ω(2k),...,ω(n*k))(4) During training, the curve features of active power time series data are extracted.

Images