EC-LSTM for Non-Intrusive Load Monitoring: A Total Energy-Conserving Approach

Non-intrusive load monitoring identifies appliance operational states by analyzing aggregate power features, yet traditional methods suffer from accuracy degradation under multi-appliance operation due to feature overlap. To address this, we propose a novel load decomposition algorithm based on Energy-Conserving Long Short-Term Memory (EC-LSTM) networks. Our methodology extends LSTM's temporal memory capabilities with conservation-driven feature learning, establishing complex nonlinear mappings between total power and individual appliance consumption. The proposed model is rigorously assessed against four baseline architectures (LSTM, GRU, Transformer, and Seq2Point) across four key metrics: Fl-score, mean absolute error (MAE), accuracy, and the Residual Energy Transfer Error (RE TE). Experimental results demonstrate that EC-LSTM achieves state-of-the-art performance, outperforming all baseline models in every metric.