A load decomposition method based on causal convolution

Abstract

The present invention belongs to the technical field of non-invasive load decomposition, and specifically is a load decomposition method based on causal convolution. The present invention processes four signal sequences of current, active power, reactive power and apparent power, and inputs them into a neural network with causal convolution as the main component. The network model predicts the probability distribution of the current input sequence based on all previously generated data, thereby generating a corresponding signal sequence. Experiments have shown that the best generation result should be a current sequence, which is used as an electrical appliance fingerprint decomposed from aggregated data loads and can be used to determine the electrical appliance operation-related information contained in the input signal sequence. One of the biggest advantages of this method is that it can realize multiple input methods, such as current, active power, and a combination of active power and reactive power, to help improve the accuracy of load decomposition.

Description

Technical Field

[0001] The present invention belongs to the technical field of non-intrusive load decomposition, and in particular is a load decomposition method based on causal convolution. Background Art

[0002] Electricity meters can only report the total household power load data, so how to infer the power load of specific equipment and perform power load decomposition is the main research goal at this stage. How to use deep learning-related methods to improve the accuracy of load decomposition is a hot topic in current research. Some successful methods, including hidden Markov models (HMM), cross entropy methods, and integer programming, do not have causal relationships and cannot be used for real-time applications. In addition, many methods only use voltage or active power as input data for load decomposition, without using data such as current, reactive power, and apparent power. There is still a lot of room for improvement. Summary of the invention

[0003] The purpose of the present invention is to provide a load decomposition method based on causal convolution to solve the problems of poor real-time performance and incomplete data utilization.

[0004] The load decomposition method based on causal convolution provided by the present invention processes four signal sequences of current, active power, reactive power and apparent power, and inputs them into a neural network with causal convolution as the main component. The network model predicts the probability distribution of the current input sequence based on all the data generated before, thereby generating a corresponding signal sequence. Experiments have shown that the best generation result should be a current sequence, which is used as an electrical appliance fingerprint decomposed from the aggregated data load, and can be used to determine the electrical appliance operation related information contained in the input signal sequence. One of the biggest advantages of this method is that it can realize multiple input methods (such as current, active power, and active power and reactive power combination input methods) to help improve the accuracy of load decomposition, thereby solving the problems of poor real-time performance and incomplete data utilization in the above-mentioned technology.

[0005] The load decomposition method based on causal convolution provided by the present invention comprises the following specific steps:

[0006] Step 1: Process the meter record data and determine the input data;

[0007] Step 2: Construct a gated version of the dilated causal convolution;

[0008] Step 3: Determine the overall network structure;

[0009] Step 4: Determine the evaluation indicators and train the network model;

[0010] Step 5: The signal sequence generated by the network is used as the appliance fingerprint for aggregated data load decomposition to obtain information related to appliance operation.

[0011] Each step is further described below.

[0012] Step 1: Process the meter record data to determine the input data

[0013] First of all, it is clear that electrical appliance fingerprints are mainly divided into two aspects:

[0014] ① High-frequency sampling: mainly manifested in the state switching (On / Off) between the steady state and transient state of the electrical appliance, as well as the signal waveform under continuous load state;

[0015] ② Low-frequency sampling: mainly manifested as root mean square (RMS), that is, current I, active power P, reactive power Q, apparent power S and other related data.

[0016] Step 1.1: This paper proposes a load decomposition method for non-intrusive load monitoring (NILM), which is to decompose the aggregated data recorded from the central meter (such as a household smart meter) into a single measurement value to estimate the usage of different load devices. The main recorded data includes two types, active power P and reactive power Q. In order to obtain all the low-frequency sampling related data, the following mathematical calculation formula is required:

[0017] S = I·V;

[0018] P = S·cos(θ);

[0019] Q = S sin(θ);

[0020] Wherein, V represents voltage, I represents current, and θ represents the phase of voltage relative to current (i.e., phase angle).

[0021] Step 1.2: This method can realize multiple data input modes, such as inputting current I, active power P separately, or inputting the combined data of active power P and reactive power Q, ensuring that the data are sampled at the same time step and adding the combined data.

[0022] Step 2 builds a gated version of the dilated causal convolution, which includes:

[0023] Step 2.1: Causal convolution predicts the probability distribution of the current sample based on all the samples that have been generated before; let the length of the input data be M samples (one sample is sampled every 1 minute), so the probability of generating a sample p(x) within time T is:

[0024]

[0025] Step 2.2: Because the input data is long, the dilated convolution can expand the receptive field. The dilated convolution input is x[m], the dilation factor is D, and the parameter c is M. k , so the output y[m] is:

[0026]

[0027] Although the input length is M, the receptive field of the output y is D·(M-1)+1. By stacking dilated convolutions and continuously improving the dilation factor, it is possible to maintain causality, sampling rate, and the use of all input characteristics while greatly improving the receptive field.

[0028] Step 2.3: The gated version is implemented by two activation functions. The output of each convolutional layer enters the sigmoid activation function and the modified linear activation function at the same time. The output of the two functions is multiplied as the output of the current module. At the same time, the output is divided into two parts, one part is used as the input of the next layer, and the other part uses stride connection, skipping all subsequent convolutions and passing it directly to the last layer of the network. In addition, the dropout of each layer is 10%.

[0029] Step 3: Determine the overall network structure

[0030] The overall structure of the network consists of the following three parts:

[0031] Fully connected layer: The fully connected layer contains M time-distributed nodes. Each time node is independent of each other. All input data is input to the convolutional layer module through the fully connected layer in strict time order.

[0032] Convolutional layer module: includes 9 gated versions of dilated convolutional layers, each layer contains a different number of filters, that is, has a specific dilation factor. At the same time, the output of this module is divided into two parts, one part is output to the next layer of the network, and the other part is connected to the last layer of the network through the residual module;

[0033] Output layer: This layer first includes a fully connected layer to ensure that the output signal length corresponds to the input. When using aggregated data input, the output of this layer also passes through a tanh activation function. The mask output by this function ensures that the entire network only outputs the type of data that we want to decompose.

[0034] Step 4: Determine the evaluation indicators and train the network model

[0035] Step 4.1: Because the network of this method is a regression network, load decomposition is achieved by generating relevant data. Therefore, the evaluation criterion during network training should be the error between the generated data and the real data, using the general evaluation notation "estimation accuracy" (S. Makonin and F. Popowich, "Nonintrusive load monitoring (NILM) performance evaluation," Energy Efficiency, vol. 8, no. 4, pp. 809–814, 2015):

[0036]

[0037] in, is the data generated by appliance k at time t, s k (t) is the real data; T represents the total time, and K is the total number of electrical appliances.

[0038] In some special cases, such as for a specific appliance, the total calculation may not be performed. This is expressed in Est.Acc. as removing the sum calculation for k and only calculating the estimation accuracy for the estimation result of the specific device, and using the estimation accuracy to calculate the loss value. The input data length is M. To ensure causality, the data is overlapped during training, and the loss value is calculated from M-1 to the total data length N so that all available data can be used for training at the same time.

[0039] Step 5: Generate signal sequence using network model

[0040] The trained network in step 4 only generates specific signals, mainly current data. This method realizes the decomposition of the fingerprint data of the target electrical appliance in two situations. One is to regard the signals from other electrical loads in the aggregated data as measurement noise, and decompose the target data from them; the other is to decompose the fingerprint of the appliance from the real load data aggregated by the specific electrical appliance (same time length, different measurement conditions).

[0041] Compared with the prior art, the present invention has the following beneficial effects:

[0042] The present invention processes four signal sequences of current, active power, reactive power and apparent power, and inputs them into a neural network with causal convolution as the main component. The network model predicts the probability distribution of the current input sequence based on all the data generated before, thereby generating a corresponding signal sequence. Experiments have shown that the best generation result should be the current sequence, which is used as the electrical appliance fingerprint decomposed from the aggregated data load and can be used to determine the electrical appliance operation related information contained in the input signal sequence. One of the biggest advantages of this method is that it can realize multiple input methods (such as current, active power, and active power and reactive power combination input methods) to help improve the accuracy of load decomposition. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 The present invention is a flow chart of a load decomposition method based on causal convolution. DETAILED DESCRIPTION

[0044] The technical solution of the present invention is described in detail below in conjunction with the accompanying drawings and embodiments.

[0045] The present invention proposes a load decomposition method based on causal convolution, the flow chart of which is as follows: Figure 1 As shown, it can be divided into the following steps:

[0046] Step 1: Process the meter record data and determine the input data;

[0047] Step 2: Construct a gated version of the dilated causal convolution;

[0048] Step 3: Determine the overall network structure;

[0049] Step 4: Determine the evaluation indicators and train the network model;

[0050] Step 5: The signal sequence generated by the network is used as the appliance fingerprint for aggregated data load decomposition to obtain information related to appliance operation.

[0051] Each step is further described below.

[0052] Step 1. Process the meter record data and determine the input data

[0053] First of all, it is clear that electrical appliance fingerprints are mainly divided into two aspects:

[0054] ① High-frequency sampling: mainly manifested in the state switching (On / Off) between the steady state and transient state of the electrical appliance, as well as the signal waveform under continuous load state;

[0055] ② Low-frequency sampling: mainly manifested as root mean square (RMS), namely current I, active power P, reactive power Q, and apparent power S and other related data;

[0056] Step 1.1: This paper proposes a load decomposition method for non-intrusive load monitoring (NILM), which is to decompose the aggregated data recorded from the central meter (such as a household smart meter) into a single measurement value to estimate the usage of different load devices. The main recorded data includes two types, active power P and reactive power Q. In order to obtain all the low-frequency sampling related data, the following mathematical calculation formula is required:

[0057] S=I·V

[0058] P=S·cos(θ)

[0059] Q=S·sin(θ)

[0060] Where V is voltage, I is current, and θ is the phase of voltage relative to current (i.e., phase angle).

[0061] Step 1.2: This method can realize multiple data input modes, such as inputting current I, active power P separately, or inputting the combined data of active power P and reactive power Q, ensuring that the data are sampled at the same time step and adding the combined data.

[0062] Step 2. Build a gated version of the dilated causal convolution

[0063] Step 2.1: Causal convolution predicts the probability distribution of the current sample based on all the samples that have been generated before. The length of the input data is M samples (one sample is sampled every 1 minute). In this example, M is 512, so the probability of generating a sample p(x) within time T is:

[0064]

[0065] Step 2.2: Because the input data is long, the dilated convolution can expand the receptive field. The dilated convolution input is x[m], the dilation factor is D, and the parameter c is M. k , so the output y[m] is:

[0066]

[0067] Although the input length is M, the receptive field of the output y is D·(M-1)+1. By stacking dilated convolutions and continuously improving the dilation factor, it is possible to maintain causality, sampling rate, and the use of all input characteristics while greatly improving the receptive field.

[0068] Step 2.3: The gated version is implemented by two activation functions. The output of each convolutional layer enters the sigmoid activation function and the modified linear activation function at the same time. The output of the two functions is multiplied as the output of the current module. At the same time, the output is divided into two parts, one part is used as the input of the next layer, and the other part uses stride connection, skipping all subsequent convolutions and passing it directly to the last layer of the network. In addition, the dropout of each layer is 10%.

[0069] Step 3. Determine the overall network structure

[0070] The overall structure of the network consists of the following three parts:

[0071] Fully connected layer: The fully connected layer contains M time-distributed nodes. Each time node is independent of each other. All input data is input to the convolutional layer module through the fully connected layer in strict time order.

[0072] Convolutional layer module: This module mainly includes 9 gated versions of dilated convolutional layers. Each layer contains a specific number of filters (specific dilation factors), which are 512, 256, 256, 128, 128, 256, 256, 256, 512. At the same time, the output of this module is divided into two parts, one part is output to the next layer of the network, and the other part is connected to the last layer of the network through the residual module;

[0073] Output layer: This layer first includes a fully connected layer to ensure that the output signal length corresponds to the input. When using aggregated data input, the output of this layer also passes through a tanh activation function. The mask output by this function ensures that the entire network only outputs the type of data that we want to decompose.

[0074] Step 4. Determine the evaluation indicators and train the network model

[0075] Step 4.1: Because the network of this method is a regression network, load decomposition is achieved by generating relevant data. Therefore, the evaluation criterion during network training should be the error between the generated data and the real data, using the general evaluation notation "estimation accuracy":

[0076]

[0077] in is the data generated by appliance k at time t, s k (t) is the real data; T represents the total time, and K is the total number of electrical appliances.

[0078] Step 4.2: In some special cases, such as for a specific appliance, the sum calculation may not be performed. The representation in Est.Acc. is to remove the sum calculation for k, and only calculate the estimation accuracy for the estimation result of the specific device, and calculate the loss value with the estimation accuracy; the input data length is M. To ensure the causal relationship, the data will be overlapped during training, and the loss value is calculated from 512 to the total data length of 1440, so that all available data can be used for training at the same time.

[0079] Step 5. The signal sequence generated by the network is used as the appliance fingerprint for aggregated data load decomposition to obtain information related to appliance operation

[0080] In step 4, the trained network only generates current data. In this case, the signals from other electrical loads in the aggregated data are regarded as measurement noise, and the fingerprint data of the target electrical appliance is decomposed, from which the fingerprint of the target electrical appliance is decomposed. At the same time, in this embodiment, different inputs of the network are compared, and the existing load decomposition method SSHMM (S. Makonin, F. Popowich, IV Baji′c, B. Gill, and L. Bartram, "Exploiting HMMsparsity to perform online real-time nonintrusive load monitoring," IEEE Transactions on Smart Grid, vol. 7, no. 6, pp. 2575–2585, 2016.) is evaluated on the same data set. As shown in Table 1, when the input is multi-type data, more data features can be obtained, and the network can achieve better load decomposition results. When the output is current, the decomposition effect of the network is better than that of the output is power.

[0081] Table 1 Comparison of load decomposition results

[0082] Decomposition method enter Output Est.Acc. Methods in this case I I 85.6％ Methods in this case I, P, Q, S P 88.4％ Methods in this case I, P, Q, S I 94.7％ SSHMM I, P, Q, S I 90.2％ .

Claims

1. A load decomposition method based on causal convolution, characterized in that: The specific steps are: Step 1, processing the meter record data to determine the input data; First, determine the appliance fingerprint in two aspects: ① High-frequency sampling: mainly manifested in the state switching between steady state and transient state of the electrical appliance, as well as the signal waveform under continuous load state; ② Low-frequency sampling: mainly manifested as root mean square value (RMS), namely current I, active power P, reactive power Q and apparent power S related data; The aggregated data from the central meter records are broken down into single measurements to estimate the usage of different load devices; the main recorded data includes two types, active power P and reactive power Q. All low-frequency sampling related data are calculated using the following mathematical formula: S = I·V; P = S·cos(θ); Q = S sin(θ); Where V is voltage, I is current, and θ is the phase of voltage relative to current; For multiple input data modes, including inputting current I, active power P, or inputting active power P and reactive power Q combined data, ensure that the data are sampled at the same time step and add the combined data; Step 2: Construct a gated version of the dilated causal convolution, which includes: Step 2.1, causal convolution predicts the probability distribution of the current sample based on all the samples that have been generated before; let the length of the input data be M samples, and the probability of generating a sample p(x) within T time is: Step 2.2: Since the input data is long, dilated convolution is used to expand the receptive field. The dilated convolution input is x[m], the dilation factor is D, and the parameter c is M. k , the output y[m] is: Although the input length is M, the receptive field of the output y is D·(M-1)+1. By stacking dilated convolutions and continuously increasing the dilation factor, we can achieve the characteristics of maintaining causality, sampling rate, and using all inputs while greatly increasing the receptive field. Step 2.3: The gated version is implemented by two activation functions. The output of each convolutional layer enters the sigmoid activation function and the modified linear activation function at the same time. The output of the two functions is multiplied as the output of the current module. At the same time, the output is divided into two parts, one as the input of the next layer, and the other part uses stride connection, skipping all subsequent convolutions and passing directly to the last layer of the network. In addition, the dropout of each layer is 10%; Step 3: Determine the overall network structure The overall structure of the network consists of the following three parts: fully connected layer, convolutional layer module, and output layer; among them: The fully connected layer contains M nodes distributed in time, each time node is independent of each other, and all input data are input to the convolutional layer module through the fully connected layer in strict time order; The convolutional layer module includes 9 gated versions of dilated convolutional layers, each layer contains a different number of filters, that is, has a specific dilation factor, and the output of the module is divided into two parts, one part is output to the next layer of the network, and the other part is connected to the last layer of the network through the residual module; The output layer includes a fully connected layer to ensure that the output signal length corresponds to the input. When using aggregated data input, the output of this layer also passes through a tanh activation function. The mask output by this function ensures that the entire network only outputs the type of data that is expected to be decomposed; Step 4: Determine the evaluation indicators and train the network model Since the network is a regression network, load decomposition is achieved by generating relevant data; the evaluation criterion during network training is the error between the generated data and the real data, and the general evaluation notation "estimation accuracy" is used: in, is the data generated by appliance k at time t, s k (t) is the real data; T represents the total time, K is the total number of electrical appliances; For a specific appliance, the sum is not calculated, that is, the sum calculation for k is removed from the Est.Acc. formula, and only the estimation accuracy is calculated for the estimation result of the specific device, and the loss value is calculated based on the estimation accuracy; the input data length is M. To ensure the causal relationship, the data is overlapped during training, and the loss value is calculated from M-1 to the total data length N, so that all available data can be used for training at the same time; Step 5: Generate signal sequence using network model The trained network in step 4 generates a specific signal - current data, and decomposes the fingerprint data of the target appliance in two situations. One is to regard the signals from other electrical loads in the aggregated data as measurement noise and decompose the target data from them; the other is to decompose the fingerprint of the appliance from the real load data aggregated by the specific appliance.

Images