Low-frequency signal non-intrusive load state identification method based on cross-layer network connection

Abstract

The application discloses a low-frequency signal non-intrusive load state identification method based on cross-layer network connection, extracts the current, active power and reactive power of a total electric meter and respective change amounts, feeds into a defined deep neural network 1 to obtain the change amounts of the current, active power and reactive power of each single electric appliance, feeds into a defined deep neural network 2 together with the electric quantity of the total circuit connected by layer hopping to obtain the working state of the single electric appliance. The method has the advantages that: since layer hopping is adopted, high-level features and low-level features are used for reasoning at the same time, and data is fully utilized; since low-frequency sampling data and a double-layer LSTM model are used, the model structure is not complex, the operation amount is controllable, and the requirement for hardware is not high; and since the low-frequency non-intrusive signal processing technology based on 1-minute sampling is used, the method does not need to make hardware modification on the electric meter, cost is greatly saved, and the method has wide application and promotion prospects.

Description

Technical Field

[0001] This invention belongs to the field of non-intrusive load monitoring, specifically relating to a non-intrusive load status identification method based on low-frequency signals from cross-layer network connections. Background Technology

[0002] Non-intrusive load monitoring (NILM) is a method of breaking down end-user electricity consumption data into various electrical loads using smart meters. NILM can monitor each load in a household and extract its energy consumption using only smart meter data, without the use of any sensors or intervention. Studies have shown that providing consumers with device-level energy consumption feedback can save up to 20% on energy consumption per residence. NILM also helps energy providers understand their customers' consumption behavior and provides them with opportunities to improve customer satisfaction. Therefore, both providers and end-users can benefit from the meaningful information provided by NILM. Intrusive load monitoring (ILM), on the other hand, achieves load monitoring through individual sensors, but is more expensive because it requires the use of numerous sensors to collect data and transmit it to a data center, as well as the maintenance and updating of all these components. Although the first research on NILM was conducted in 1992, progress has been slow and difficult. However, with the widespread use of smart meters, accelerating research into smart homes and energy efficiency, NILM is gaining more attention. The paper "[A hierarchical hidden Markov model framework for home appliance modeling]" uses a hierarchical Hidden Markov Model (HMM) Bayesian network to model appliances, which outperforms some robust HMM variants and produces more accurate results. Although some HMM models have also achieved good results, their biggest drawback is that computational complexity increases exponentially with the number of devices, and each appliance requires detailed modeling, which is a time-consuming process and limits their practical use. Summary of the Invention

[0003] The technical problem this invention aims to solve is to address the shortcomings of existing technologies by providing a non-intrusive load status identification method based on low-frequency signals using cross-layer network connections. This method can obtain the load status of each individual appliance based on data collected from the main electricity meter, facilitating further processing. By employing layer-skipping connections and utilizing both high-level and low-level features for inference, it fully utilizes the data. The use of low-frequency sampling data and a two-layer LSTM model results in a simple model structure, controllable computational load, and low hardware requirements. Furthermore, based on low-frequency non-intrusive signal processing technology with sampling once per minute, this invention requires no hardware modification to the electricity meter, significantly reducing costs and demonstrating broad application and promotion prospects.

[0004] The technical solution adopted by this invention to solve the above-mentioned technical problems is: a non-intrusive load state identification method based on low-frequency signals of cross-layer network connections, comprising the following steps:

[0005] 1) Collect training data on the current, active power and reactive power of the main meter, and collect training data on the current, active power and reactive power of each individual appliance. Cluster the current, active power and reactive power of each individual appliance to obtain the working state set of each appliance.

[0006] 2) Using the current, active power, and reactive power of the total meter at each moment in the training data as the sequence of electrical quantities of the total circuit, calculate the changes in the total meter current, active power, and reactive power at each moment in the training data relative to the total meter current, active power, and reactive power at the previous moment to obtain the sequence of electrical changes in the total circuit; calculate the changes in the current, active power, and reactive power of each individual appliance at each moment in the training data relative to the current, active power, and reactive power of each individual appliance at the previous moment to obtain the sequence of electrical changes in each individual appliance.

[0007] 3) Construct a deep neural network 1. For each single electrical appliance, take the total circuit electrical quantity sequence and the total circuit electrical change sequence as input features, and the single electrical appliance electrical change sequence as output features. Feed them into the deep neural network 1 for learning to obtain the single electrical appliance change regression model.

[0008] 4) Construct a second deep neural network. For each single electrical appliance, take the total circuit electrical quantity sequence, the total circuit electrical change sequence, and the output of the single electrical appliance change regression model as input features, and take the single electrical appliance working state set as output features. Feed them into the second deep neural network for learning to obtain a single electrical appliance working state classification model.

[0009] 5) For each individual appliance, the total circuit electrical quantity sequence and the total circuit electrical change sequence are used as input features and fed into the individual appliance change regression model to derive the electrical change of each individual appliance; for each individual appliance, the total circuit electrical quantity sequence, the total circuit electrical change sequence, and the output of the individual appliance change regression model are used as input features, and the individual appliance operating state set is used as output features and fed into the individual appliance operating state classification model to identify the load state of each individual appliance.

[0010] As a preferred option, the specific process of step 1) is as follows:

[0011] 1.1) Given N electrical appliances, collect training data on their current, active power, and reactive power at time t. The sequences of current, active power, and reactive power in the training data are as follows: Cluster them into M classes using the mean shift clustering method, where M≤10;

[0012] When the number of clusters M > 10, the bandwidth parameter of the meanshift method is increased by 5 each time until the number of clusters M ≤ 10.

[0013] 1.2) Sort the cluster centers according to the magnitude of the current, with smaller currents first and larger currents last. Number them, and consider the group with the smallest number as the electrical appliance in the off state, otherwise it is in the on state, thus obtaining the working state set of a single electrical appliance.

[0014] As a preferred option, the specific process of step 2) is as follows:

[0015] 2.1) Sample once per minute to obtain the total current of the main meter at time t in the training data. Total active power Total reactive power As a feature of the training feature vector of the total electricity meter at time t, the total current of the total electricity meter at time t-1 in the training data is sampled. Total active power Total reactive power As features of the training feature vector of the total electricity meter at time t-1, the relative changes in the total current, active power, and reactive power of the total electricity meter at times t and t-1 are calculated, denoted as...

[0016] 2.2) Sample once per minute to obtain the current of a single electrical appliance at time t in the training data. Total active power Total reactive power As a feature of the training feature vector of a single electrical appliance at time t, the current of a single electrical appliance at time t-1 in the training data is sampled. Total active power Total reactive power As the features of the training feature vector of the single electrical appliance at time t-1, the relative changes in current, active power, and reactive power of the single electrical appliance at time t and time t-1 are calculated and denoted as .

[0017] As a preferred option, the specific process of step 3) is as follows:

[0018] 3.1) Construct a deep neural network consisting of a sequence input layer 1, an LSTM layer 1, an LSTM layer 2, a fully connected layer 1, and a regression layer 1;

[0019] 3.2) For each individual electrical appliance, sequence the total electrical quantities of the circuit. Total circuit electrical change sequence As input features, the electrical change sequence of a single electrical appliance is used. The output features are fed into a deep neural network to learn and obtain a regression model of the change in a single electrical appliance.

[0020] As a preferred option, the specific process of step 4) is as follows:

[0021] 4.1) Construct a deep neural network II consisting of a second sequence input layer, a third LSTM layer, a fourth LSTM layer, a second fully connected layer, and a second regression layer;

[0022] 4.2) For each individual electrical appliance, sequence the total electrical quantities of the circuit. Total circuit electrical change sequence The output of the single appliance change regression model is used as the input feature, and the set of working states of the single appliance is used as the output feature. These are fed into a deep neural network II for learning to obtain a single appliance working state classification model.

[0023] As a preferred option, the specific process of step 5) is as follows:

[0024] 5.1) For each individual electrical appliance, sequence the total electrical quantities of the circuit. Total circuit electrical change sequence As input features, they are fed into the single appliance change regression model to derive the electrical change of each single appliance.

[0025] 5.2) For each individual electrical appliance, sequence the total electrical quantities of the circuit. Total circuit electrical change sequence The output of the single appliance change regression model is used as the input feature, and the set of operating states of the single appliance is used as the output feature. These are then fed into the single appliance operating state classification model to identify the load state of each single appliance.

[0026] Compared with the prior art, the present invention has the following advantages:

[0027] This invention discloses a non-intrusive load state identification method based on low-frequency signals from a cross-layer network connection. By extracting the current, active power, and reactive power of the main meter and their respective changes, the method feeds them into a predefined deep neural network to obtain the changes in the current, active power, and reactive power of each individual appliance. These changes, along with the total electrical quantities of the cross-layer connected circuit, are then fed into a predefined deep neural network to obtain the operating state of each appliance.

[0028] This invention presents a non-intrusive load status identification method based on low-frequency signals from cross-layer network connections. It can determine the load status of each individual appliance based on data collected from the main electricity meter, facilitating further processing. By employing layer-skipping connections and utilizing both high-level and low-level features for inference, it fully utilizes the data. The use of low-frequency sampling data and a two-layer LSTM model results in a simple model structure, manageable computation, and low hardware requirements. Furthermore, based on low-frequency non-intrusive signal processing technology with sampling once per minute, this method requires no hardware modification to the electricity meter, significantly reducing costs and demonstrating broad application and promotion prospects. Attached Figure Description

[0029] Figure 1 The diagram shows the state of the dishwasher and its corresponding current in the embodiment.

[0030] Figure 2 This is the current diagram for the main electricity meter;

[0031] Figure 3 This is a diagram showing the actual current of the dishwasher. Detailed Implementation

[0032] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0033] The embodiment of the low-frequency signal non-intrusive load state identification method based on cross-layer network connection includes the following steps:

[0034] 1) Collect training data on the current, active power and reactive power of the main meter, and collect training data on the current, active power and reactive power of each individual appliance. Cluster the current, active power and reactive power of each individual appliance to obtain the working state set of each appliance.

[0035] The specific process of step 1) is as follows:

[0036] 1.1) Suppose there are N=6 electrical appliances, namely a basement light, dishwasher, washing machine, refrigerator, television, and computer. Collect training data on their current, active power, and reactive power at time t. The current, active power, and reactive power sequences in the training data are as follows: Cluster them into M classes using the mean shift clustering method, where M≤10;

[0037] When the number of clusters M > 10, the bandwidth parameter of the meanshift method is increased by 5 each time until the number of clusters M ≤ 10.

[0038] 1.2) Sort the cluster centers according to the magnitude of the current, with smaller currents first and larger currents last. Number them, and consider the cluster with the smallest number as the electrical appliance in the off state, otherwise it is in the on state, thus obtaining the working state set of a single electrical appliance.

[0039] 2) Using the current, active power, and reactive power of the total meter at each moment in the training data as the sequence of electrical quantities of the total circuit, calculate the changes in the total meter current, active power, and reactive power at each moment in the training data relative to the total meter current, active power, and reactive power at the previous moment to obtain the sequence of electrical changes in the total circuit; calculate the changes in the current, active power, and reactive power of each individual appliance at each moment in the training data relative to the current, active power, and reactive power of each individual appliance at the previous moment to obtain the sequence of electrical changes in each individual appliance.

[0040] The specific process of step 2) is as follows:

[0041] 2.1) Sample once per minute to obtain the total current of the main meter at time t in the training data. Total active power Total reactive power As a feature of the training feature vector of the total electricity meter at time t, the total current of the total electricity meter at time t-1 in the training data is sampled. Total active power Total reactive power As features of the training feature vector of the total electricity meter at time t-1, the relative changes in the total current, active power, and reactive power of the total electricity meter at times t and t-1 are calculated, denoted as...

[0042] 2.2) Sample once per minute to obtain the current of a single electrical appliance at time t in the training data. Total active power Total reactive power As a feature of the training feature vector of a single electrical appliance at time t, the current of a single electrical appliance at time t-1 in the training data is sampled. Total active power Total reactive power As the features of the training feature vector of the single electrical appliance at time t-1, the relative changes in current, active power, and reactive power of the single electrical appliance at time t and time t-1 are calculated and denoted as .

[0043] 3) Construct a deep neural network 1. For each single electrical appliance, take the total circuit electrical quantity sequence and the total circuit electrical change sequence as input features, and the single electrical appliance electrical change sequence as output features. Feed them into the deep neural network 1 for learning to obtain the single electrical appliance change regression model.

[0044] The specific process of step 3) is as follows:

[0045] 3.1) Construct a deep neural network consisting of a sequence input layer 1, an LSTM layer 1, an LSTM layer 2, a fully connected layer 1, and a regression layer 1;

[0046] 3.2) For each individual electrical appliance, sequence the total electrical quantities of the circuit. Total circuit electrical change sequence As input features, the electrical change sequence of a single electrical appliance is used. As the output feature, it is fed into a deep neural network to learn and obtain a regression model of the change in a single electrical appliance.

[0047] 4) Construct a second deep neural network. For each single electrical appliance, take the total circuit electrical quantity sequence, the total circuit electrical change sequence, and the output of the single electrical appliance change regression model as input features, and take the single electrical appliance working state set as output features. Feed them into the second deep neural network for learning to obtain a single electrical appliance working state classification model.

[0048] The specific process of step 4) is as follows:

[0049] 4.1) Construct a deep neural network II consisting of a second sequence input layer, a third LSTM layer, a fourth LSTM layer, a second fully connected layer, and a second regression layer;

[0050] 4.2) For each individual electrical appliance, sequence the total electrical quantities of the circuit. Total circuit electrical change sequence The output of the single appliance change regression model is used as the input feature, and the set of working states of the single appliance is used as the output feature. These are fed into a deep neural network II for learning to obtain a single appliance working state classification model.

[0051] 5) For each individual appliance, the total circuit electrical quantity sequence and the total circuit electrical change sequence are used as input features and fed into the individual appliance change regression model to derive the electrical change of each individual appliance; for each individual appliance, the total circuit electrical quantity sequence, the total circuit electrical change sequence, and the output of the individual appliance change regression model are used as input features, and the operating state set of the individual appliance is used as output features and fed into the individual appliance operating state classification model to identify the load state of each individual appliance;

[0052] The specific process of step 5) is as follows:

[0053] 5.1) For each individual electrical appliance, sequence the total electrical quantities of the circuit. Total circuit electrical change sequence As input features, they are fed into the single appliance change regression model to derive the electrical change of each single appliance.

[0054] 5.2) For each individual electrical appliance, sequence the total electrical quantities of the circuit. Total circuit electrical change sequence The output of the single appliance change regression model is used as the input feature, and the set of operating states of the single appliance is used as the output feature. These are then fed into the single appliance operating state classification model to identify the load state of each single appliance.

[0055] In this embodiment, the dishwasher load status identification results are shown below. Figures 1-3 . Figure 1 The diagram shows the state of the dishwasher and its corresponding current in the embodiment. Figure 2 This is the current diagram for the main electricity meter. Figure 3 This is the actual current diagram of the dishwasher, expressed in terms of the current at the corresponding cluster centers.

Claims

1. A non-intrusive load condition identification method for low-frequency signals based on cross-layer network connections, characterized in that, Includes the following steps: 1) Collect training data on the current, active power and reactive power of the main meter, and collect training data on the current, active power and reactive power of each individual appliance. Cluster the current, active power and reactive power of each individual appliance to obtain the working state set of each appliance. 2) Using the current, active power, and reactive power of the total meter at each moment in the training data as the sequence of electrical quantities of the total circuit, calculate the changes in the total meter current, active power, and reactive power at each moment in the training data relative to the total meter current, active power, and reactive power at the previous moment to obtain the sequence of electrical changes in the total circuit; calculate the changes in the current, active power, and reactive power of each individual appliance at each moment in the training data relative to the current, active power, and reactive power of each individual appliance at the previous moment to obtain the sequence of electrical changes in each individual appliance. 3) Construct a deep neural network 1. For each single electrical appliance, take the total circuit electrical quantity sequence and the total circuit electrical change sequence as input features, and the single electrical appliance electrical change sequence as output features. Feed them into the deep neural network 1 for learning to obtain the single electrical appliance change regression model. 4) Construct a second deep neural network. For each single electrical appliance, take the total circuit electrical quantity sequence, the total circuit electrical change sequence, and the output of the single electrical appliance change regression model as input features, and take the single electrical appliance working state set as output features. Feed them into the second deep neural network for learning to obtain a single electrical appliance working state classification model. 5) For each individual appliance, the total circuit electrical quantity sequence and the total circuit electrical change sequence are used as input features and fed into the individual appliance change regression model to derive the electrical change of each individual appliance; for each individual appliance, the total circuit electrical quantity sequence, the total circuit electrical change sequence, and the output of the individual appliance change regression model are used as input features, and the individual appliance operating state set is used as output features and fed into the individual appliance operating state classification model to identify the load state of each individual appliance.

2. The method for non-intrusive load state identification of low-frequency signals based on cross-layer network connections according to claim 1, characterized in that, The specific process of step 1) is as follows: 1.1) Given N electrical appliances, collect training data on their current, active power, and reactive power at time t. The sequences of current, active power, and reactive power in the training data are as follows: k′=1,2,...N, use the meanshift clustering method to cluster them into M classes, requiring M≤10; When the number of clusters M > 10, the bandwidth parameter of the meanshift method is increased by 5 each time until the number of clusters M ≤ 10. 1.2) Sort the cluster centers according to the magnitude of the current, with smaller currents first and larger currents last. Number them, and consider the group with the smallest number as the electrical appliance in the off state, otherwise it is in the on state, thus obtaining the working state set of a single electrical appliance.

3. The non-intrusive load state identification method for low-frequency signals based on cross-layer network connections according to claim 2, characterized in that, The specific process of step 2) is as follows: 2.1) Sample once per minute to obtain the total current of the main meter at time t in the training data. Total active power Total reactive power As a feature of the training feature vector of the total electricity meter at time t, the total current of the total electricity meter at time t-1 in the training data is sampled. Total active power Total reactive power As features of the training feature vector of the total electricity meter at time t-1, the relative changes in the total current, active power, and reactive power of the total electricity meter at times t and t-1 are calculated, denoted as... 2.2) Sample once per minute to obtain the current of a single electrical appliance at time t in the training data. Total active power Total reactive power As a feature of the training feature vector of a single electrical appliance at time t, the current of a single electrical appliance at time t-1 in the training data is sampled. Total active power Total reactive power As the features of the training feature vector of the single electrical appliance at time t-1, the relative changes in current, active power, and reactive power of the single electrical appliance at time t and time t-1 are calculated and denoted as .

4. The non-intrusive load state identification method for low-frequency signals based on cross-layer network connections according to claim 3, characterized in that, The specific process of step 3) is as follows: 3.1) Construct a deep neural network consisting of a sequence input layer 1, an LSTM layer 1, an LSTM layer 2, a fully connected layer 1, and a regression layer 1; 3.2) For each individual electrical appliance, sequence the total electrical quantities of the circuit. Total circuit electrical change sequence As input features, the electrical change sequence of a single electrical appliance is used. The output features are fed into a deep neural network to learn and obtain a regression model of the change in a single electrical appliance.

5. The non-intrusive load state identification method for low-frequency signals based on cross-layer network connections according to claim 4, characterized in that, The specific process of step 4) is as follows: 4.1) Construct a deep neural network II consisting of a second sequence input layer, a third LSTM layer, a fourth LSTM layer, a second fully connected layer, and a second regression layer; 4.2) For each individual electrical appliance, sequence the total electrical quantities of the circuit. Total circuit electrical change sequence The output of the single appliance change regression model is used as the input feature, and the set of working states of the single appliance is used as the output feature. These are fed into a deep neural network II for learning to obtain a single appliance working state classification model.

6. The method for non-intrusive load state identification of low-frequency signals based on cross-layer network connections according to claim 5, characterized in that, The specific process of step 5) is as follows: 5.1) For each individual electrical appliance, sequence the total electrical quantities of the circuit. Total circuit electrical change sequence As input features, they are fed into the single appliance change regression model to derive the electrical change of each single appliance. 5.2) For each individual electrical appliance, sequence the total electrical quantities of the circuit. Total circuit electrical change sequence The output of the single appliance change regression model is used as the input feature, and the set of operating states of the single appliance is used as the output feature. These are then fed into the single appliance operating state classification model to identify the load state of each single appliance.

Images