Evaluating causal CNN variants for NILM: reproducing and extending WaveNILM

Non-intrusive load monitoring (NILM) is the task of estimating appliance-level energy consumption from aggregate electrical signals. The resulting information enables detailed feedback on device usage, supporting behavioral changes and informed decision-making for improved energy efficiency through data-driven monitoring and management of electrical systems. Recent approaches, such as WaveNILM, employ causal one-dimensional convolutional neural networks to capture long-range temporal dependencies. This study introduces a modified dilated architecture that integrates two techniques not previously applied in this context: separable one-dimensional convolutions, which decrease computational cost, and channel-wise Squeeze-and-Excitation (cSE) attention, which highlights informative features. Batch normalization is included to enhance training stability. The model is evaluated on the AMPds2 dataset with noisy aggregate inputs. Experiments considered both full disaggregation of all monitored appliances and focused detection of five flexible-use appliances, using four types of input signals: current, active power, reactive power, and apparent power. The proposed model achieved higher accuracy than WaveNILM in three of the four input cases, with the largest gains for apparent power. In addition, the number of trainable parameters was reduced by 31.2%, demonstrating a more efficient and scalable design.