The multi-scale feature extraction method effectively captures a wealth of feature information

while the attention mechanism filters out significant load characteristics. Multi-task learning enhances the utilization of inter-task correlations. Building on these advantages

we propose a non-intrusive load monitoring approach based on parallel multi-scale feature extraction

aimed at more accurately identifying and analyzing the energy consumption patterns of electrical appliances. The proposed method employs a multi-scale convolutional network to capture multi-dimensional load features across various time scales. It integrates an attention mechanism to selectively highlight important features from multiple perspectives. By leveraging a multi-task learning strategy

the model enhances the complementary relationship between state estimation and load decomposition

improving overall performance. Experiments conducted on the UK-DALE dataset reveal that the proposed model achieves superior performance in load disaggregation and appliance state estimation

with an average on/off state recognition accuracy exceeding 97%

and lower disaggregation errors compared to benchmark models. Furthermore

the model's effectiveness and generalizability are confirmed through real-world sampling data

particularly in the precise detection of charging power for electric bicycles

which assists in the timely identification of unsafe residential charging practices and contributes to fire prevention efforts.