Volume 60

An Attention-Enhanced Deep Learning Method Tailored for Non-IntrusiveLoad Monitoring in Air Conditioning Systems Chuanyu Tang, Yi Gong, Qiong Chen, Nan Li

Abstract

The increasing energy demand for air conditioningsystems presents substantial economic andenvironmental challenges, making effective airconditioning load monitoring crucial. Traditionalintrusive load monitoring techniques were costly andchallenging to implement, whereas non-intrusive loadmonitoring (NILM) technology offered a cost-effectivealternative. This study constructed the CN-AC-NILMdataset, customized for the characteristics of Chineseelectrical appliances and cultural usage patterns, andsubsequently developed an RNN-attention model for airconditioning systems, with a focus on evaluating theimpact of the attention mechanism on the model’sperformance. Initially, the optimal input feature set forthe air conditioning load disaggregation predictionmodel was determined using the forward featureselection method based on the dataset, followed by anevaluation of performance differences among variousRNN models. Subsequently, it was observed that in thecomparative analysis of baseline models applied to airconditioning system for NILM, LSTM exhibited higheraccuracy than GRU, and bidirectional modelsoutperformed unidirectional models. Furthermore,after introducing the attention mechanism in thebaseline models, the average accuracy of RNN modelsincreased by 21.09%, accompanied by a 5.33% increasein average training time. Finally, the effects of attentionmechanisms and bidirectional structure on the modelswere compared, revealing that the introduction of theattention mechanism outperformed the bidirectionalstructure, and the LSTM-Attention exhibited the bestoverall performance for NILM of the air conditioningsystems. Notably, Attention mechanisms enhancedmodel interpretability through the visualization ofattention weights. By improving load disaggregationaccuracy and interpretability, this attention-enhancedmodel enables more precise energy management for airconditioning systems. It reduces electricity costs,optimizes energy use, and supports demand-sidemanagement, offering substantial economic savings andlowering environmental impact, especially in highenergy-demand settings.

Keywords Attention mechanism, Recurrent neural network, Non-intrusive load monitoring, Air conditioning system, CN-AC-NILM dataset