LSR-NILM: A Non-Invasive Load Monitoring Method Fusing Timing and Time-Frequency Features

Non-Intrusive Load Monitoring (NILM) is an efficient energy management tool to monitor the usage of home appliances by obtaining fine-grained power usage information at the power inlet and analyzing the power consumption data. This study aims to investigate the improvement methods of NILM system to enhance the accuracy and adaptability of load recognition and proposes a deep learning based NILM architecture LSTM-SE-ResNet for NILM (LSR-NILM). In this study, time series data are segmented by sliding window technique and features are extracted from time domain and time-frequency domain. The time-domain features and current envelope statistical analysis are used to capture the operating characteristics of the load, while the time-frequency features are used to refine the energy distribution in different frequency bands through wavelet packet decomposition, and the main harmonic components are extracted by combining with the Fourier transform to form a complete set of features. Finally, these features are fed into the LSR-NILM architecture, which combines Convolutional LSTM (ConvLSTM) for temporal feature learning, and introduces Squeeze-and-Excitation Block (SE-Block) in the Attention Mechanism to strengthen the weight of key features, and adopts the Residual Network (ResNet). ResNet is used to improve the efficiency of feature transfer and reduce the gradient vanishing problem. Experimental results on the open dataset Plug-Load Appliance Identification Dataset (PLAID) show that the proposed method outperforms the existing methods in terms of accuracy (94.83%) compared to existing methods (88.13%~94.72%), achieving higher load identification accuracy while maintaining lower model complexity, demonstrating its effectiveness and practicality in NILM applications.