Scaled and Inter-token Relation Enhanced Transformer for Sample-restricted Residential NILM

Recent advancements in transformer models have yielded impressive results in Non-Intrusive Load Monitoring (NILM). However, effectively training a transformer on small-scale datasets remains a challenge. This paper addresses this issue by enhancing the attention mechanism of the original transformer to improve performance. We propose two novel mechanisms: the inter-token relation enhancement mechanism and the dynamic temperature tuning mechanism. The first mechanism reduces the prioritization of intra-token relationships in the token similarity matrix during training, thereby increasing inter-token focus. The second mechanism introduces a learnable temperature tuning for the token similarity matrix, mitigating the over-smoothing problem associated with fixed temperature values. Both mechanisms are supported by rigorous mathematical foundations. We evaluate our approach using the REDD residential NILM dataset, a relatively small-scale dataset and demonstrate that our methodology significantly enhances the performance of the original transformer model across multiple appliance types.