A non-intrusive load monitoring method based on a residual fully convolutional neural network

Abstract

The present invention discloses a non-intrusive load monitoring method based on a residual fully convolutional neural network. First, the total power and load data of each device are collected and data preprocessing operations are performed; then, in an adaptive sliding window manner, feature engineering extraction is carried out on the time series data within each window to obtain corresponding feature data; subsequently, a residual fully convolutional neural network model is established, a loss function is constructed, and the network parameters are trained; then, a power feature database is established for the target device, and the actual activation sequence decomposition value is compared with the activation sequence features to eliminate irrelevant activations generated by the network; finally, an evaluation model is established to evaluate the accuracy of the output time series, thus completing non-intrusive load monitoring.

Description

Technical Field

[0001] The present invention belongs to the technical field of non - intrusive load monitoring, and mainly relates to a non - intrusive load monitoring method based on a residual fully convolutional neural network. Background Art

[0002] The core idea of non - intrusive load monitoring (NILM) is to separate the aggregated load data into individual device - level power sequences by analyzing the operating characteristics of each device. According to the sampling rate of the load data, NILM methods can be divided into two categories: low - frequency - based methods and high - frequency - based methods. High - frequency - based NILM methods are more effective in detecting transient events and differentiating appliances that exhibit similar power consumption profiles due to more detailed appliance characteristics. However, the high computational requirements and the cost of obtaining high - frequency sampling data can pose significant challenges for real - time implementation. Therefore, choosing the appropriate NILM method for a specific application requires a trade - off between the accuracy of device - level information and the feasibility of implementation. Recent research has shown an increasing interest in low - frequency - based NILM algorithms due to the widespread use of smart meters.

[0003] However, load decomposition based on low - frequency data introduces a series of difficulties that need to be addressed to ensure reliable results. One important challenge is the limited temporal resolution of low - frequency measurements. Since these measurements are taken over larger time intervals, they may not capture rapid load fluctuations and changes, which are common in some types of loads. This limitation can lead to ambiguity in load patterns, affecting the accurate identification of individual sub - loads. In addition, low - frequency data tends to mask transient load changes and short - term load events, which are crucial for accurate load decomposition. Rapid transitions caused by devices such as refrigerators, air conditioners, or elevators may be lost in the coarse - grained low - frequency data, thus hindering the ability to accurately distinguish between different sub - loads. As a result, the final decomposition model may overlook these key load dynamics, leading to a loss of fine - grainedness in the results.

[0004] At the same time, the inherent noise problem in low - frequency data also constitutes an important obstacle. Since these measurements cover a wider time span, they are more susceptible to cumulative noise, measurement errors, and other interferences. These noises may introduce inaccuracies in the decomposition process, especially during periods of low - intensity loads or low overall power consumption. These inaccuracies not only affect the quality of the decomposition results but also the ability to effectively identify the contributions of minor loads. To address these challenges, a comprehensive approach is needed that integrates more advanced noise reduction techniques to improve the quality of load decomposition. Summary of the Invention

[0005] In view of the deficiencies in the prior art, the present invention provides a non-intrusive load monitoring method based on a residual fully convolutional neural network. First, the total power and load data of each device are collected and data preprocessing operations are performed; then, in an adaptive sliding window manner, feature engineering extraction is performed on the time series data within each window to obtain corresponding feature data; subsequently, a residual fully convolutional neural network model is established, a loss function is constructed, and the network parameters are trained; then, a power feature database is established for the target device, and the actual activation sequence decomposition value is compared with the activation sequence features to eliminate irrelevant activations generated by the network; finally, an evaluation model is established to evaluate the accuracy of the output time series, completing non-intrusive load monitoring.

[0006] To achieve the above object, the technical solution adopted by the present invention is: a non-intrusive load monitoring method based on a residual fully convolutional neural network, comprising the following steps:

[0007] S1, data preprocessing: Collect the total power and load data of each device and perform data preprocessing operations;

[0008] S2, data segmentation: Based on load feature extraction with an adaptive window length, using the data preprocessed in step S1, in an adaptive sliding window manner, feature engineering extraction is performed on the time series data within each window to obtain corresponding feature data; the data is divided into a training set and a test set and input into the residual fully convolutional neural network;

[0009] S3, establish a residual fully convolutional neural network model: The residual fully convolutional neural network model includes six residual convolutional modules, three one-dimensional convolutional layers, four maxpooling layers, four ConvTranspose layers, a flatten layer, and a dense layer; each residual convolutional module consists of four one-dimensional convolutional layers. Among them, the input of the first one-dimensional convolutional layer in each residual convolutional module is added to the output of the fourth one-dimensional convolutional layer through a residual connection. Each one-dimensional convolutional layer is composed of 30 convolutional kernels with a dimension of 8 and uses a ReLU activation function; the training set time, total power consumption data, and single-device power consumption data are input into the residual fully convolutional neural network module;

[0010] S4, construct a loss function and train the network parameters: By inputting the total power consumption data of the test set, using the residual fully convolutional neural network model established in step S3, the corresponding single-electrical appliance time series data is decomposed;

[0011] S5, posterior processing: Establish a power feature database for the target device, and compare the actual activation sequence decomposition value with the activation sequence features to eliminate irrelevant activations generated by the network;

[0012] S6. Establish an evaluation model: Evaluate the accuracy of the output time series.

[0013] As an improvement of the present invention, the data preprocessing in step S1 specifically includes the following steps:

[0014] S11. Data cleaning and filling: Receive historical and real-time power load data, perform data cleaning operations, delete or correct outliers and missing data; among them, for missing data, use interpolation for filling; for abnormal data, use the Z-score method outlier detection algorithm to identify and process outliers;

[0015] S12. Feature selection and extraction: Select key features from the original power load data and perform statistical information extraction; the key features at least include time, date, and season; the statistical information at least includes mean and variance;

[0016] S13. Standardization and normalization: Perform standardization and normalization processing on the features selected in step S12, and select min-max normalization to ensure that features of different scales can be treated equally in the model;

[0017] S14. Time series smoothing: For power load data with obvious seasonality and periodicity, use the time series smoothing method.

[0018] As an improvement of the present invention, in step S2, the collected load sequence is segmented by the sliding window method based on the adaptive window length, and different sampling rates and window lengths are applied for different loads; the window length is defined as follows:

[0019] W(i) = [Wbase * f_t(i) * T(i)] / [fbase * Tbase]

[0020] Where, W(i) represents the window length calculated for the i-th type of device, T(i) is the average working cycle of the i-th type of device, f_t(i) is the sampling frequency of the i-th type of device, and Wbase, fbase, and Tbase respectively represent the base window length, base sampling frequency, and base working cycle.

[0021] As another improvement of the present invention, in the residual fully convolutional neural network model in step S3: the output of the maxpooling layer 1 is connected to the input of the residual convolution module 1; the output of the residual convolution module 1 is connected to the input of the maxpooling layer 2; the output of the maxpooling layer 2 is connected to the input of the residual convolution module 2; the output of the residual convolution module 2 is connected to the input of the maxpooling layer 3; the output of the maxpooling layer 3 is connected to the input of the residual convolution module 3; the output of the residual convolution module 3 is connected to the input of the maxpooling layer 4; the output of the maxpooling layer 4 is connected to three one-dimensional convolutional layers; the output of the three one-dimensional convolutional layers is connected to the input of the ConvTranspose layer 1; the output of the ConvTranspose layer 1 is connected to the input of the residual convolution module 4; the output of the residual convolution module 4 is connected to the input of the ConvTranspose layer 2; the output of the ConvTranspose layer 2 is connected to the input of the residual convolution module 5; the output of the residual convolution module 5 is connected to the input of the ConvTranspose layer 3; the output of the ConvTranspose layer 3 is connected to the input of the residual convolution module 6; the output of the residual convolution module 6 is connected to the input of the ConvTranspose layer 4; the output of the ConvTranspose layer 4 is connected to the input of the flatten layer; the output of the flatten layer is connected to the input of the dense layer; the output of the maxpooling layer 1 is added to the input of the ConvTranspose layer 4 through a residual connection; the output of the maxpooling layer 2 is added to the input of the ConvTranspose layer 3 through a residual connection; the output of the maxpooling layer 3 is added to the input of the ConvTranspose layer 2 through a residual connection; the output of the maxpooling layer 4 is added to the input of the ConvTranspose layer 1 through a residual connection.

[0022] As another improvement of the present invention, the loss function in step S4 is as follows:

[0023]

[0024] Wherein, T represents the total number of sampling points included in each sample; N represents the total number of samples; p gt and p pd respectively represent the actual power value and the device power time series obtained by network decomposition.

[0025] As yet another improvement of the present invention, step S5 specifically includes:

[0026] S51, set the data filtering threshold:

[0027] S511: Select appropriate data samples, and select a group of samples that do not appear in the training and test sets from historical and real-time data;

[0028] S512: Determine the switch state of the device. For a device with a clear switch state, use the integrated function in NILMTK to determine the switching moment of each device type on the selected data;

[0029] S513: Process the data of the continuous output system. For a continuous output system, through similarity calculation, adopt the device threshold with a similar energy scale;

[0030] S52. Analyze the characteristics of the actual power sequence: Calculate the actual power characteristics. Based on the set activation threshold, calculate the actual power sequence characteristics of the target device at each activation, and record the actual power sequence values and the associated characteristics of each activation;

[0031] S53. Analyze the activation length and operations: Calculate the activation length of each target device each time by calculating the difference between timestamps; Determine the minimum activation length in the operations by recording the activation sequence of the target device, and use the total number of sampling points of the activation sequence as the threshold;

[0032] S54. Eliminate irrelevant decomposition results: Obtain the activation information of the power of the target device obtained by the algorithm decomposition, compare the minimum activation length of the actual value with the activation length of the decomposition value, and eliminate all activations with power less than the threshold and activations with a duration less than the minimum activation length in the list from the decomposition value.

[0033] As a further improvement of the present invention, the evaluation indexes of the evaluation model in step S6 include the mean absolute error, the normalized signal aggregation error, and the normalized decomposition error, and the calculation of the indexes is as follows:

[0034]

[0035]

[0036] Among them, MAE represents the mean absolute error; NDE represents the normalized signal aggregation error; SAE represents the normalized decomposition error; M represents the total number of sampling points; p gt and p pd are respectively the actual power value and the decomposed power value of the device.

[0037] Compared with the prior art, the beneficial effects of the present invention are as follows: The present invention provides a non-intrusive load monitoring method based on a residual fully convolutional neural network. Based on the operating cycles and characteristics of different devices, an adaptive window length data segmentation method is adopted to ensure the effectiveness of identifying device behaviors. A residual fully convolutional neural network model structure is also proposed for the energy decomposition of household appliances and distributed energy. By using residual connections, the problems of model performance degradation and gradient disappearance that occur with the increase in network depth are successfully solved, improving the algorithm accuracy and generalization ability. In addition, the method of the present invention also has a posterior processing algorithm to optimize the output of the deep neural network. Considering the data characteristics, a power feature database is established for the target device, and the actual activation sequence decomposition value is compared with the activation sequence features to eliminate irrelevant activations generated by the network. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 FIG. is a flowchart of the steps of the non-intrusive load monitoring method based on a residual fully convolutional neural network of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0039] The present invention will be further clarified below in conjunction with the drawings and specific embodiments. It should be understood that the following specific embodiments are only used to illustrate the present invention and not to limit the scope of the present invention.

[0040] Embodiment 1

[0041] A non-intrusive load monitoring method based on a residual fully convolutional neural network includes the following steps, and the implementation block diagram is as Figure 1 shown:

[0042] Step S1: Perform data preprocessing on the original data. The specific steps are as follows:

[0043] Step S11: Data cleaning and filling. Receive historical and real-time power load data, perform data cleaning operations to delete or correct outliers and missing data; for missing data, use interpolation for filling to ensure data integrity and accuracy; for abnormal data, use the Z-score method outlier detection algorithm to identify and process possible outliers;

[0044] Step S12: Feature selection and extraction. Select key features from the original power load data, including time, date, season, etc., and extract statistical information, such as mean, variance, etc., to reduce the data dimension and retain important information;

[0045] Step S13: Standardization and normalization. Perform standardization and normalization processing on the selected features, and select min-max normalization to ensure that features of different scales can be equally treated in the model;

[0046] Step S14: Time series smoothing. For power load data with obvious seasonality and periodicity, apply time series smoothing techniques such as moving average or exponential smoothing to reduce noise and sudden fluctuations and improve the stability of model prediction.

[0047] Step S2: Data segmentation. Based on load feature extraction with an adaptive window length, use the preprocessed data and adopt an adaptive sliding window method to perform feature engineering extraction on the time series data within each window to obtain corresponding feature data; divide the data into a training set and a test set and input them into the residual fully convolutional neural network.

[0048] Use a sliding window method based on an adaptive window length to segment the collected load sequences, and apply different sampling rates and window lengths for different load settings. The window length is defined as follows:

[0049] W(i) = [Wbase * f_t(i) * T(i)] / [fbase * Tbase]

[0050] where W(i) represents the window length calculated for the i-th type of device, T(i) is the average working cycle of the i-th type of device, f_t(i) is the sampling frequency of the i-th type of device, and Wbase, fbase, and Tbase represent the base window length, base sampling frequency, and base working cycle respectively.

[0051] Step S3: Establish a residual fully convolutional neural network model, and its specific structure includes:

[0052] Six residual convolutional modules, three one-dimensional convolutional layers, four maxpooling layers, four ConvTranspose layers, a flatten layer, and a dense layer; the structure of the residual convolutional module is as follows. Each residual convolutional module consists of four one-dimensional convolutional layers, each one-dimensional convolutional layer is composed of 30 convolutional kernels with a dimension of 8, and all use the ReLU activation function; in each residual convolutional module, the input of the first one-dimensional convolutional layer is added to the output of the fourth one-dimensional convolutional layer through a residual connection;

[0053] The connection method of each module of the residual fully convolutional neural network model is as follows:

[0054] The output of the maxpooling layer 1 is connected to the input of the residual convolution module 1; the output of the residual convolution module 1 is connected to the input of the maxpooling layer 2; the output of the maxpooling layer 2 is connected to the input of the residual convolution module 2; the output of the residual convolution module 2 is connected to the input of the maxpooling layer 3; the output of the maxpooling layer 3 is connected to the input of the residual convolution module 3; the output of the residual convolution module 3 is connected to the input of the maxpooling layer 4; the output of the maxpooling layer 4 is connected to three one-dimensional convolution layers; the outputs of the three one-dimensional convolution layers are connected to the input of the ConvTranspose layer 1; the output of the ConvTranspose layer 1 is connected to the input of the residual convolution module 4; the output of the residual convolution module 4 is connected to the input of the ConvTranspose layer 2; the output of the ConvTranspose layer 2 is connected to the input of the residual convolution module 5; the output of the residual convolution module 5 is connected to the input of the ConvTranspose layer 3; the output of the ConvTranspose layer 3 is connected to the input of the residual convolution module 6; the output of the residual convolution module 6 is connected to the input of the ConvTranspose layer 4; the output of the ConvTranspose layer 4 is connected to the input of the flatten layer; the output of the flatten layer is connected to the input of the dense layer; the output of the maxpooling layer 1 is added to the input of the ConvTranspose layer 4 through a residual connection; the output of the maxpooling layer 2 is added to the input of the ConvTranspose layer 3 through a residual connection; the output of the maxpooling layer 3 is added to the input of the ConvTranspose layer 2 through a residual connection; the output of the maxpooling layer 4 is added to the input of the ConvTranspose layer 1 through a residual connection.

[0055] Step S4, construct a loss function, and the loss function is as follows:

[0056] Train the network by minimizing the loss function, and use the following function as the loss function in the training of the deep neural network, which is defined as:

[0057]

[0058] where, T represents the total number of sampling points included in each sample; N represents the total number of samples; p gt and p pd respectively represent the actual power value and the device power time series obtained by network decomposition.

[0059] Step S5: Posterior processing, construct a data posterior processing method, compare the decomposed power sequence with the operation feature database established using the actual ground truth value, which can eliminate the irrelevant activations generated by the deep neural network. The specific steps are as follows:

[0060] Step S51: Set the data filtering threshold

[0061] First, select appropriate data samples. Select a set of samples from historical and real-time data that do not appear in the training and test sets. These samples will be used for subsequent analysis and adjustment.

[0062] Second, determine the switch state devices: For devices with a clear switch state, use the integrated functions in NILMTK to determine the switching moments of each device type on the selected data. Process the data of continuous output systems: For continuous output systems, through similarity calculation, use the threshold of devices with a similar energy scale.

[0063] Step S52: Analyze the characteristics of the actual power sequence

[0064] First, calculate the actual power characteristics: Based on the set activation threshold, calculate the characteristics of the actual power sequence of the target device at each activation. Include the timestamps of the start and stop of operation, the maximum and average power, etc. Then, record the actual power sequence values and the associated characteristics of each activation for use in subsequent analysis and adjustment.

[0065] Step S53: Analyze the activation length and operations

[0066] Calculate the activation length: Calculate the length for each activation of the target device by calculating the difference between timestamps. This helps to understand the running duration and activation frequency of the device.

[0067] Determine the minimum activation length: By recording the activation sequence of the target device, determine the minimum activation length in the operation, and use the total number of sampling points of the activation sequence as the threshold.

[0068] Step S54: Eliminate irrelevant decomposition results

[0069] Obtain the activation information of the power of the target device obtained by the algorithm decomposition, and compare the minimum activation length of the actual value with the activation length of the decomposition value. Eliminate all activations with power less than the threshold and activations with a duration less than the minimum activation length of the activations in the list from the decomposition value.

[0070] Step S6: Establish an evaluation model to evaluate the decomposition accuracy. The evaluation model includes the following metrics:

[0071] Use the mean absolute error, normalized signal aggregation error, and normalized decomposition error. The calculation formulas are

[0072]

[0073]

[0074]

[0075] Among them, MAE represents the mean absolute error; NDE represents the normalized signal aggregation error; SAE represents the normalized decomposition error; M represents the total number of sampling points; p gt and p pd are the actual power value and the decomposed power value of the device, respectively.

[0076] Test case

[0077] The decomposition results are tested by using the non-intrusive load monitoring method based on the residual fully convolutional neural network proposed in the present invention as follows:

[0078] In this test case, the combined dataset of the public datasets REDD and Pecan Street is used as the experimental data. House 1, 3, 4, and 5 in the REDD dataset are used as the training set, and house 2 and 6 are used as the test set. In the Pecan Street dataset, the load data collected from six independent households in the Austin area in the same month are used as the test set and the training set respectively. The trained model decomposes the load data of the test set, and the evaluation model established is used to evaluate the generated results. The obtained indicators are shown in Table 1:

[0079] Table 1 Comparison table of the decomposition accuracy of the load decomposition method based on the convolutional neural network and the present method

[0080]

[0081]

[0082] As can be seen from Table 1, compared with the load decomposition method based on the convolutional neural network, the decomposition accuracy of the three types of devices by the present method has been significantly improved. The model indicators of the present method are all significantly better than those of the load decomposition method based on the convolutional neural network, indicating the accuracy of the model in decomposing the new power system. Among them, the MAEs of electric vehicles, refrigerators, and microwave ovens have decreased by 54.86%, 69.18%, 58.85%, and 75.08% respectively. Among them, compared with the load decomposition method based on the convolutional neural network, the decomposition value accuracy of the microwave oven by the present method has the largest improvement. This may be because the operating cycle of the microwave oven is short and the overall working time ratio is small, so its feature extraction is more vulnerable to the influence of photovoltaic power generation.

[0083] At the same time, compared with the method of directly outputting without posterior processing, it can be seen that by eliminating the irrelevant activations in the model, the method of posterior correction further reduces the decomposition error and improves the comprehensive decomposition accuracy of the model.

[0084] It should be noted that the above content only illustrates the technical idea of the present invention and cannot be used to limit the protection scope of the present invention. For those of ordinary skill in the art, without departing from the principle of the present invention, several improvements and refinements can be made, and these improvements and refinements all fall within the protection scope of the claims of the present invention.

Claims

1. A non-intrusive load monitoring method based on a residual fully convolutional neural network, characterized in that , it includes the following steps: S1, Data preprocessing: Collect the total power and load data of each device, and perform data preprocessing operations; S2, Data segmentation: Based on load feature extraction with an adaptive window length, use the data preprocessed in step S1, and adopt an adaptive sliding window method to extract feature engineering for the time series data within each window to obtain corresponding feature data; Divide the data into a training set and a test set, and input them into the residual fully convolutional neural network; Among them, a sliding window method based on an adaptive window length is used to segment the collected load sequence, and different sampling rates and window lengths are applied for different loads; The window length is defined as follows: W(i) = [Wbase * f_t(i) * T(i)] / [fbase * Tbase] where W(i) represents the window length calculated for the i-th type of device, T(i) is the average working cycle of the i-th type of device, f_t(i) is the sampling frequency of the i-th type of device, and Wbase, fbase, and Tbase represent the base window length, base sampling frequency, and base working cycle respectively; S3, Establish a residual fully convolutional neural network model: The residual fully convolutional neural network model includes six residual convolutional modules, three one-dimensional convolutional layers, four maxpooling layers, four ConvTranspose layers, a flatten layer, and a dense layer; Each residual convolutional module consists of four one-dimensional convolutional layers. Among them, the input of the first one-dimensional convolutional layer in each residual convolutional module is added to the output of the fourth one-dimensional convolutional layer through a residual connection. Each one-dimensional convolutional layer consists of 30 convolutional kernels with a dimension of 8 and uses the ReLU activation function; Input the training set time, total power consumption data, and single-device power consumption data into the residual fully convolutional neural network module; S4, Construct a loss function and train the network parameters: By inputting the total power consumption data of the test set, use the residual fully convolutional neural network model established in step S3 to decompose the corresponding single-electrical appliance time series data; S5, Posterior processing: Establish a power feature database for the target device, and compare the actual activation sequence decomposition value with the activation sequence feature to eliminate irrelevant activations generated by the network; S6, Establish an evaluation model: Evaluate the accuracy of the output time series.

2. A non-intrusive load monitoring method based on a residual fully convolutional neural network according to claim 1, characterized in that: The data preprocessing in step S1 specifically includes the following steps: S11, Data cleaning and filling: Receive historical and real-time power load data, perform data cleaning operations, delete or correct outliers and missing data; Among them, for missing data, interpolation is used for filling; For abnormal data, the Z-score method outlier detection algorithm is used to identify and process outliers; S12, Feature Selection and Extraction: Select key features from the original power load data and perform statistical information extraction; the key features at least include time, date, and season; the statistical information at least includes mean and variance; S13, Standardization and Normalization: Perform standardization and normalization processing on the features selected in step S12, and choose min-max normalization to ensure that features of different scales can be equally treated in the model; S14, Time Series Smoothing: For power load data with obvious seasonality and periodicity, use time series smoothing methods.

3. A non-intrusive load monitoring method based on a residual fully convolutional neural network as described in claim 2, characterized in that: In the residual fully convolutional neural network model in step S3: The output of maxpooling layer 1 is connected to the input of residual convolution module 1; the output of residual convolution module 1 is connected to the input of maxpooling layer 2; the output of maxpooling layer 2 is connected to the input of residual convolution module 2; the output of residual convolution module 2 is connected to the input of maxpooling layer 3; the output of maxpooling layer 3 is connected to the input of residual convolution module 3; the output of residual convolution module 3 is connected to the input of maxpooling layer 4; the output of maxpooling layer 4 is connected to three one-dimensional convolutional layers; the outputs of the three one-dimensional convolutional layers are connected to the input of ConvTranspose layer 1; the output of ConvTranspose layer 1 is connected to the input of residual convolution module 4; the output of residual convolution module 4 is connected to the input of ConvTranspose layer 2; the output of ConvTranspose layer 2 is connected to the input of residual convolution module 5; the output of residual convolution module 5 is connected to the input of ConvTranspose layer 3; the output of ConvTranspose layer 3 is connected to the input of residual convolution module 6; the output of residual convolution module 6 is connected to the input of ConvTranspose layer 4; the output of ConvTranspose layer 4 is connected to the input of the flatten layer; the output of the flatten layer is connected to the input of the dense layer; the output of maxpooling layer 1 is added to the input of ConvTranspose layer 4 through a residual connection; the output of maxpooling layer 2 is added to the input of ConvTranspose layer 3 through a residual connection; the output of maxpooling layer 3 is added to the input of ConvTranspose layer 2 through a residual connection; the output of maxpooling layer 4 is added to the input of ConvTranspose layer 1 through a residual connection.

4. A non-intrusive load monitoring method based on a residual fully convolutional neural network as described in claim 3, characterized in that: The loss function in step S4 is as follows: where T represents the total number of sampling points included in each sample; N represents the total number of samples; p gt and p pd respectively represent the actual power value and the device power time series obtained by network decomposition.

5. A non-intrusive load monitoring method based on a residual fully convolutional neural network as described in claim 4, characterized in that: Step S5 specifically includes: S51, Set the data filtering threshold: S511: Select appropriate data samples, and select a set of samples from historical and real-time data that do not appear in the training and test sets; S512: Determine the switch status devices. For devices with a clear switch status, use the integrated functions in NILMTK to determine the switching moments of each device type on the selected data; S513: Process the data of continuous output systems. For continuous output systems, through similarity calculation, adopt the device thresholds with similar energy scales; S52, Analyze the characteristics of the actual power sequence: Calculate the actual power characteristics. Based on the set activation threshold, calculate the actual power sequence characteristics of the target device at each activation, and record the actual power sequence values and the associated characteristics of each activation; S53, Analyze the activation length and operations: Calculate the activation length of each target device each time by calculating the difference between timestamps; By recording the activation sequence of the target device, determine the minimum activation length in the operations, and use the total number of sampling points of the activation sequence as the threshold; S54, Eliminate irrelevant decomposition results: Obtain the activation information of the power of the target device obtained by algorithm decomposition, compare the minimum activation length of the actual value with the activation length of the decomposition value, and eliminate all activations with power less than the threshold and activations with a duration less than the minimum activation length of the activations in the list in the decomposition value.

6. A non-intrusive load monitoring method based on a residual fully convolutional neural network as claimed in claim 5, characterized in that: The evaluation metrics for evaluating the model in step S6 include the mean absolute error, the normalized signal aggregation error, and the normalized decomposition error, and the calculation of the metrics is as follows: Among them, MAE represents the mean absolute error; NDE represents the normalized signal aggregation error; SAE represents the normalized decomposition error; M represents the total number of sampling points; p gt and p pd are the actual power value and the decomposed power value of the device, respectively.

Images