ChronoFuse-TCN: A progressive temporal convolutional network for multi-scale and spatiotemporal load disaggregation

Abstract

To address the limitations of existing Non-Intrusive Load Monitoring (NILM) methods in capturing the multi-scale variability and spatiotemporal dependencies of appliance power consumption, this paper proposes a novel progressive temporal convolutional architecture, ChronoFuse-TCN. The proposed model adopts a multi-stage feature extraction strategy to progressively enhances its ability to represent and interpret appliance-level load patterns. By combining dynamic multi-scale modeling, long-range temporal context encoding, and spatiotemporal attention mechanisms, the proposed approach enables more effective separation of overlapping and dynamic power signals. Furthermore, cross-stage feature integration is employed to enrich the hierarchical representation of load features. Experimental results on the UK-DALE dataset show that ChronoFuse-TCN achieves significantly lower disaggregation error compared to state-of-the-art baselines, demonstrating its effectiveness and generalization capability in complex NILM scenarios.