DeepANFIS: an interpretable and uncertainty-aware neuro-fuzzy framework for energy disaggregation in smart microgrids

Non-intrusive load monitoring (NILM) facilitates appliance-level energy disambiguation based on aggregate power consumption measurements and is an important facilitator of energy efficiency and demand-side management. However, state-of-the-art NILM methods tend to have low interpretability, low sensitivity to the sampling rate, and difficulty in providing reliable uncertainty information, which delay their effectiveness in practical implementations. To address these issues, we present a new deep adaptive neuro-fuzzy inference system (DeepANFIS) for joint appliance-state identification and power estimation with enhanced interpretability and uncertainty awareness. The proposed system comprehensively incorporates modular convolutional feature learning with multiple resolutions, fuzzy memory networks with temporal dependencies, hierarchical fuzzy inference, and evidential deep learning. Extensive experiments were conducted on both the UK-DALE dataset and REDD to illustrate the efficacy of DeepANFIS under high-frequency and low-frequency sampling schemes. On the UK-DALE dataset, DeepANFIS attained an F1-score of 0.7122 with a mean absolute error (MAE) of 25.63 W at a sampling frequency of 6 s, along with an F1-score of 0.8371 and a root mean-squared error (RMSE) of 78.03 W at a sampling frequency of 15 min. On the REDD, DeepANFIS obtained a near-perfect F1-score of 0.9869 with an accuracy of 0.9992 and a low MAE of 2.56 W. As part of the cross-dataset analysis, evaluation on REDD showed excellent F1-score classification capability with a slight drop of 2.4%, but the power estimation transfer remained a challenge. An ablation analysis was further performed to validate the significances of both temporal memory and fuzzy processing in DeepANFIS. Thus, the efficiency of DeepANFIS was established with respect to a non-interpretable uncertainty model for NILM.