Deep Neural Networks (DNNs) have been the subject of much research over the years, with a particular emphasis on Convolutional Neural Networks (CNNs) and Recurrent Neural Networks (RNNs). The purpose of this study is to evaluate the energy usage of single-source appliances, like smart meters, in the context of Non-intrusive Load Monitoring (NILM).This paper introduces a novel workflow for NILM by combining a regression subnetwork with a classification subnetwork. The architecture incorporates an RNN with attention mechanism in the regression and classification networks, drawing inspiration from successful applications in temporal attention. The experiments conducted on our own dataset called SmartSense show that our proposed architecture outperforms the state-of-the-art under various conditions.