Bidirectional Attention LSTM Networks for Non-instructive Load Monitoring

Abstract

Non-instructive load monitoring (NILM) is a data processing method that decomposes the total energy consumption and estimates the power of individual electrical appliances. The application of NILM can provide additional information for optimal control strategy of smart grid, to achieve the purpose of saving energy by fine management. However, the accuracy of traditional NILM methods doesn’t have high accuracy of decomposed power value. In this work, we apply long short-term memory (LSTM) and achieve good accuracy by enhancing the LSTM model with bidirectional and attention mechanisms, as well as kernel density estimation. The model first normalizes the total energy consumption and converts the normalized data to time series of fixed length. LSTM extracts features from the time series, with the bidirectional mechanism to operate from both normal and reverse order and the attention mechanism to calculate the attention weights of different time steps. Besides, kernel density estimation is used to fit the training data and modify the output of the deep learning model, which upgrades the disaggregation accuracy. The proposed model is tested on UK-dale dataset.