A non-intrusive load monitoring system based on decoupled attention mechanism

Abstract

The application discloses a kind of non-invasive load monitoring systems based on decoupling attention mechanism, belong to electric power load monitoring technical field, solve the problem that different users' multiple loads cannot maintain higher accuracy, i.e. poor generalization performance of non-invasive load monitoring method, the present application includes data acquisition module, data preprocessing module, feature extraction module, feature processing module and feature mapping module;The data acquisition module is used to collect and transmit the power consumption data of user;Data preprocessing module includes data storage unit and data processing unit;Feature extraction module includes one-dimensional convolution unit and maximum value pooling unit;Feature processing module includes relative position coding unit and decoupling attention score unit;Feature mapping module includes one-dimensional transpose convolution unit and fully connected output unit.The present application can more accurately capture the correlation between source sequence and target sequence, and further improve the generalization performance of non-invasive load monitoring algorithm.

Description

Technical Field

[0001] This invention belongs to the field of power load monitoring technology, specifically relating to a non-intrusive load monitoring system based on a decoupled attention mechanism. Background Technology

[0002] Load monitoring can acquire operating status information of various electrical devices, providing solutions for addressing energy supply shortages, developing energy efficiency, promoting demand-side management, and managing residential user loads. Compared to intrusive load monitoring, non-intrusive load monitoring (NILM) is lower in cost, more practical, and has a broader development prospect.

[0003] In reality, although non-invasive load monitoring (NILM) is considered a more practical load monitoring method, most existing research on NILM is still limited to identifying a relatively limited range of loads. Whether NILM can achieve high-precision load identification for the diverse loads within user premises and in the market remains highly controversial. Specifically, the most commonly used method in the literature is to identify loads by pre-establishing a feature database. This involves recording the characteristics of loads in experiments, comparing the characteristics of loads in actual operation with the feature database, and selecting the load with the closest characteristics. However, feature databases lack universality, and the identification results obtained using feature databases may contain misjudgments. Different judgments may be made for different households and different scenarios, which will affect the load identification results and further affect the accuracy of load monitoring data, making it impractical in practice.

[0004] Therefore, maintaining high identification accuracy for various types of loads across different users, i.e. ensuring good generalization performance when applied to different users, is a key condition for the application and promotion of NILM. Thus, a deep learning method based on decoupled attention mechanism is needed to improve the generalization performance of non-intrusive load monitoring. Summary of the Invention

[0005] The purpose of this invention is:

[0006] To address the issue that existing non-intrusive load monitoring methods cannot maintain high accuracy when dealing with multiple types of loads from different users, i.e., they have poor generalization performance, a non-intrusive load monitoring system based on a decoupled attention mechanism is proposed. This system comprehensively considers the content information and time-series information of the power sequence when extracting abstract features of the source sequence, which can more accurately capture the correlation between the source sequence and the target sequence, thereby improving the generalization performance of the non-intrusive load monitoring algorithm.

[0007] The technical solution adopted in this invention is as follows:

[0008] A non-intrusive load monitoring system based on a decoupled attention mechanism includes a data acquisition module, a data preprocessing module, a feature extraction module, a feature processing module, and a feature mapping module;

[0009] The data acquisition module is used to collect and transmit users' electricity consumption data;

[0010] The data preprocessing module includes a data storage unit and a data processing unit. The data storage unit is used to store the electricity consumption data transmitted by the data acquisition module, and the data processing unit is used to correct abnormal data in the electricity consumption data and store the corrected electricity consumption data into the data storage unit.

[0011] The feature extraction module includes a one-dimensional convolutional unit and a max pooling unit, which are used to extract features from the power sequence.

[0012] The feature processing module includes a relative position encoding unit and a decoupled attention scoring unit. The relative position encoding unit is used to supplement the power sequence with temporal information, and the decoupled attention unit is used to calculate the attention score that takes into account both the power sequence content information and the temporal information.

[0013] The feature mapping module includes a one-dimensional transposed convolutional unit and a fully connected output unit, which are used to obtain an output result that matches the original power sequence.

[0014] Furthermore, the specific operation steps of the feature processing module are as follows:

[0015] The implementation process of the feature processing module is analyzed, and the attention score is described by the following formula:

[0016]

[0017] In the formula: Softmax is the weighting function, d k Let A represent the dimension of the source vector, and A% be the decoupling matrix that considers the correlation between power sequence content information and time series information. Its calculation process is described by the following formula:

[0018] Q c =HW q,c ,K c =HW k,c V c =HW v,c ;

[0019]

[0020] In the formula: Q c K c and V cIt is a linear mapping matrix of the reaction power sequence content information, Q r and K r It is a linear mapping matrix that reflects the time-series information of the power sequence, where H is the original power sequence and P is the relative position encoding sequence of the original power.

[0021] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are:

[0022] The non-intrusive load monitoring system based on a decoupled attention mechanism of this invention eliminates the explicit extraction process of load features. When extracting abstract features from the source sequence, this system comprehensively considers both the content information and temporal information of the power sequence, enabling it to more accurately capture the correlation between the source and target sequences, thereby improving the general applicability of the non-intrusive load monitoring algorithm. The feature processing of this invention supplements the original power sequence with temporal information through a positional encoding scheme. In particular, the relative positional encoding scheme used in this invention better reflects the inherent relationships between data compared to the absolute positional encoding scheme, and has wider applicability.

[0023] The deep learning method of this invention provides an end-to-end solution for data processing. For non-intrusive load monitoring tasks, the deep learning method hides the process of explicitly extracting load features. By decoupling the attention mechanism, it can extract deeper abstract features contained in the original data and still has a good decomposition effect in complex scenarios. Attached Figure Description

[0024] Appendix Figure 1 This is a block diagram of the load monitoring system of the present invention.

[0025] Appendix Figure 2 This is a diagram showing the unit composition of the load monitoring system of the present invention. Detailed Implementation

[0026] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0027] A non-intrusive load monitoring system based on a decoupled attention mechanism is characterized by comprising a data acquisition module, a data preprocessing module, a feature extraction module, a feature processing module, and a feature mapping module;

[0028] The data acquisition module is used to collect and transmit users' electricity consumption data;

[0029] The data preprocessing module is used to save and correct the user's electricity consumption data. It includes a data storage unit and a data processing unit. The data storage unit is used to store the electricity consumption data transmitted by the data acquisition module. The data processing unit is used to correct abnormal data in the electricity consumption data and store the corrected electricity consumption data into the data storage unit.

[0030] The feature extraction module is used to extract abstract features from the original power sequence, including one-dimensional convolutional units and max pooling units, which are used to extract features from the power sequence.

[0031] The feature processing module is used to process the abstract features of the original power sequence, including a relative position encoding unit and a decoupled attention scoring unit. The relative position encoding unit is used to supplement the power sequence with temporal information, and the decoupled attention unit is used to calculate the attention score that takes into account the power sequence content information and temporal information.

[0032] The feature mapping module is used to output decomposition results that match the original power, including a one-dimensional transposed convolutional unit and a fully connected output unit. The one-dimensional transposed convolutional unit and the fully connected output unit are used to obtain output results that match the original power sequence.

[0033] The specific operation steps of the feature processing module are as follows:

[0034] The implementation process of the feature processing module is analyzed, and the attention score is described by the following formula:

[0035]

[0036] In the formula: Softmax is the weighting function, d k Let A represent the dimension of the source vector, and A% be the decoupling matrix that considers the correlation between power sequence content information and time series information. Its calculation process is described by the following formula:

[0037] Q c =HW q,c ,K c =HW k,c V c =HW v,c ;

[0038]

[0039] In the formula: Q c K c and V c It is a linear mapping matrix of the reaction power sequence content information, Q r and K r It is a linear mapping matrix that reflects the time-series information of the power sequence, where H is the original power sequence and P is the relative position encoding sequence of the original power.

[0040] The specific implementation of the present invention is as follows:

[0041] First, electricity consumption data from users is collected using smart meters. A data processing unit corrects bad data and supplements missing data. The processed electricity consumption data is then stored in a data storage unit. In the feature extraction module, a one-dimensional convolutional layer and a max-pooling unit extract features from the original power sequence to obtain a preliminary feature vector. Simultaneously, in the feature processing module, a relative position encoding unit encodes the original power sequence to obtain a relative position encoding vector. Then, a decoupled attention scoring unit calculates an attention score that considers both the power sequence content and temporal information. Finally, in the feature mapping module, a one-dimensional transposed convolutional unit and a fully connected output unit obtain an output result that matches the original power sequence.

[0042] The overall process of this invention is as follows: user electricity consumption data is collected through smart meters, and then abstract features considering the content information and time sequence information of the power sequence are obtained through a decoupled attention mechanism. Finally, a power decomposition result matching the original power sequence is obtained.

[0043] To comprehensively verify the performance of this invention, power decomposition tests were conducted on different appliances from different households. Refrigerators, microwave ovens, washing machines, and dishwashers from two different households were selected as research subjects. These four appliances have different operating characteristics: refrigerators are appliances that operate for long periods and exhibit periodicity; microwave ovens are appliances that operate for short periods and have many state changes; washing machines and dishwashers have longer operating times and more complex characteristics.

[0044] To comprehensively evaluate the performance of this invention, multiple evaluation indicators were selected to comprehensively measure its power decomposition effect. Four indicators were chosen: Acc, F1 score, mean relative error (MRE), and mean absolute error (MAE) to comprehensively evaluate the performance of this invention. Acc measures the accuracy of the invention in predicting the power of electrical appliances when they are in operation and off states; F1 score measures the accuracy of the invention in predicting the power of electrical appliances when they are in operation; and MRE and MAE measure the accuracy of the invention in predicting the power values ​​of electrical appliances.

[0045] The power decomposition performance evaluation indicators of this invention and three comparative benchmark models are shown in the table below:

[0046] Home 1 power decomposition results

[0047]

[0048] Home 2 power decomposition results

[0049]

[0050] The results show that the present invention can achieve good power decomposition effect in complex operating environments with various households and loads, and has good versatility.

[0051] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A non-intrusive load monitoring system based on a decoupled attention mechanism, characterized in that, It includes a data acquisition module, a data preprocessing module, a feature extraction module, a feature processing module, and a feature mapping module; The data acquisition module is used to collect and transmit users' electricity consumption data; The data preprocessing module includes a data storage unit and a data processing unit. The data storage unit is used to store the electricity consumption data transmitted by the data acquisition module, and the data processing unit is used to correct abnormal data in the electricity consumption data and store the corrected electricity consumption data into the data storage unit. The feature extraction module includes a one-dimensional convolutional unit and a max pooling unit, which are used to extract features from the power sequence. The feature processing module includes a relative position encoding unit and a decoupled attention scoring unit. The relative position encoding unit is used to supplement the power sequence with temporal information, and the decoupled attention unit is used to calculate the attention score that takes into account both the power sequence content information and the temporal information. The feature mapping module includes a one-dimensional transposed convolutional unit and a fully connected output unit, which are used to obtain an output result that matches the original power sequence. The specific operation steps of the feature processing module are as follows: The implementation process of the feature processing module is analyzed, and the attention score is described by the following formula: ； In the formula: Softmax For the weight function, d k The dimension of the source vector after feature extraction. The decoupling matrix, which takes into account the correlation between power sequence content information and time sequence information, is calculated using the following formula: ； ； In the formula: Q c , K c and V c It is a linear mapping matrix of the reaction power sequence content information. Q r and K r It is a linear mapping matrix of the reaction power sequence time information. H The original power sequence, P The relative position encoding sequence for the original power.

Images