Towards Robust NILM: A Unified Seq2Point Benchmarking Framework with Balanced Model

NILM estimates appliance-level power consumption from household aggregate power data, offering insights into energy consumption without intrusive sub-metering techniques. Although deep learning-based sequence-to-point (Seq2Point) models are becoming increasingly popular, no unified framework exists for fair and reproducible evaluation across datasets and appliance types. We address this by developing a modular bench-marking pipeline, incorporating standardised data pre-processing and consistent training and evaluation procedures. The toolkit supports three major public datasets (UKDALE, REFIT, ECO) and multiple Seq2Point variants. In addition, we propose a novel architecture, Seq2Point Balanced. It integrates a variety of regularisation techniques and architectural simplifications to significantly reduce the parameter count by 94% (1.8 million vs. 30 million), whilst improving robustness and generalisation. Experiments show that Seq2Point Balanced achieves competitive and often superior performance, particularly in noisy and cross-domain scenarios. The model demonstrates strong fine-tuning capabilities. The toolkit is publicly released to support reproducibility at: https://github.com/ymahajan1612/seq2point-update.