State disaggregation for Non-Intrusive Load Monitoring (NILM) with calibrated bagged Seq2Point 1D convolutional neural networks

Non-intrusive load monitoring (NILM) aims to infer appliance operating states from aggregate power data, yet reliable state disaggregation remains challenging because appliance labels are often generated using heuristic thresholds, ON/OFF classes are highly imbalanced, prediction probabilities are poorly calibrated, and output sequences can become fragmented and unstable. To address these limitations, this research proposes a calibrated bagged Seq2Point one-dimensional convolutional neural network, termed CB-S2P NILM. The framework integrates validation-driven per-device threshold optimization, class-balanced binary cross-entropy, temperature scaling, temporal smoothing, and dwell-time constraints, together with bagged ensemble prediction to improve robustness and predictive stability. Evaluated on the REDD dataset across six appliances, the proposed method demonstrated strong and consistent performance. In the validation phase, the Bagged CNN achieved appliance-level accuracies of 0.98 to 0.99 and F1-scores of 0.98 to 0.99, outperforming the baseline CNN across all evaluated appliances. Sensitivity analysis further showed that an ensemble size of three yielded the best overall performance, with an average accuracy of 0.99 and an average F1-score of 0.99. Overall, the proposed CB-S2P NILM provides a robust, interpretable, and reproducible framework for appliance-state disaggregation within the evaluated REDD cross-house setting.