Enhanced Appliance Disaggregation Using UNet-NILM: A Comparative Study on UK-DALE and BLOND Datasets

This paper presents an improved approach to appliance disaggregation using the UNet-NILM model, evaluated on the UK-DALE and BLOND datasets. The preprocessing of the UK-DALE dataset involved handling irregular time steps and missing data by resampling to a consistent 6-second interval and applying quantile filters to smooth the signals. The BLOND dataset, although pre-summarized to one-second intervals, required additional preprocessing to identify consistent appliance usage patterns and optimize window sizes and thresholds. The performance of the UNet-NILM model, featuring 7-layer down- and upsampling blocks, was compared to a 1D-CNN baseline, demonstrating superior disaggregation accuracy across multiple appliances, particularly for complex devices like washing machines and dishwashers. Both models were trained using identical hyperparameters, with the UNet-NILM model showing significant improvements in Mean Absolute Error (MAE), Normalized Disaggregation Error (NDE), and Estimated Accuracy (EAC) over the baseline. Experiments on the BLOND dataset further validated the scalability and robustness of the UNet-NILM model when applied to a more challenging dataset with varying appliance combinations. The results underscore the potential of UNet-NILM in achieving higher disaggregation accuracy, contributing to more effective energy monitoring and management systems.