Household Non-intrusive Load Monitoring by Graph Signal Processing and Bi-LSTM

High precision non-intrusive load monitoring provides a superior solution for electricity usage monitoring. However, the sparsity of load features obstructs the improvement of non-intrusive load disaggregation (NILM) model accuracy. To this regard, a load dis-aggregation framework integrating graph signal processing (GSP) and bidirectional long short-term memory networks (Bi-LSTM) is proposed. To address the sparsity of load features, this framework utilizes a highly efficient GSP to provide direct load features for Bi-LSTM. The Bi-LSTM integrates the outputs from GSP and total power and generates precise single load power. We also designed the application process of GSP and the implementation scheme of Bi-LSTM. We compare our method with baseline methods on the UK-DALE dataset. Experiments demonstrate the excellence of our method. Compared to the baseline method Bi-LSTM, the mean square error (MAE) is improved by 12% on average.