An adaptive lifelong incremental learning framework for power quality disturbance detection under concept drift

The expeditious rise in renewable penetration, inverter-based resources, electric vehicle charging infrastructure, and evolving load composition has driven significant concept drift and new power quality disturbance (PQD) patterns in modern power grids. As a result, contemporary static classifiers become progressively ineffective. To address this challenge, this paper proposes an adaptive lifelong incremental learning framework unified with drift detection, active online learning, and physics-based interpretability for PQD detection and classification under concept drift. Within a unified closed-loop classification architecture, four incremental learning strategies are systematically investigated: Fine-Tuning, Elastic Weight Consolidation (EWC)-Only, Replay-Only, Hybrid (with two additonal configurations of Hybrid-Abrupt and Hybrid-Mixed). The Hybrid strategy combines EWC, herding-based coreset replay, knowledge distillation, prototype regularization, and dark experience replay (DER), together with a dynamically expandable classification head. These methods are evaluated on a testbed of 20 PQDs divided into four tasks. The Hybrid method achieves the maximum accuracy ( ) with a forgetting rate of just and a Macro F1 of , validated across three seeds. A comparative analysis against continual learning baselines—iCaRL, GEM, DER , ER-ACE, and Cumulative-Oracle—demonstrates the competitive performance of the Hybrid approach. For drift detection, a KNN-embedded technique is evaluated against the established ADaptive WINdowing (ADWIN) and Drift Detection Method (DDM) techniques and is found to consistently achieve superior performance under evolving disturbance conditions. The proposed models use a ResNet-MLP-based encoder backbone; to validate its strength, the Hybrid method is additionally examined using Plain-MLP, Deep-MLP, MLP-No-BN, Transformer-Encoder, CNN-1D, and CNN-LSTM encoders. Physics-aligned features—root-mean-square (RMS) voltage and total harmonic distortion (THD)—serve dual roles as interpretable monitoring metrics and drift attribution indicators. A detailed computational complexity analysis is conducted, covering time, space, computational (FLOPs), and sample/scalability complexity. The six frameworks are further assessed under 27 noise scenarios spanning 9 categories—Pink and Brownian noise, additive white Gaussian noise (AWGN), impulsive noise, load harmonics, non-stationary mixed noise, communication signal drop, environmental flicker, and measurement bias. For the first time in the PQD literature, a comprehensive bifurcation analysis identifies the regime boundaries governed by EWC weight and noise intensity , complemented by a parametric sensitivity analysis of the proposed models. All six models are further exposed to (i) grid topology and (ii) renewable and power-electronics disturbance scenarios for realistic, system-level examination. Finally, validation on the XPQRS benchmark dataset (four tasks) shows the Hybrid-Mixed model achieving the best accuracy of 85.75 % with a forgetting rate of just 1.50 %. The complete framework was further validated through Processor-in-the-Loop (PIL) testing on a Raspberry Pi 5, confirming functional correctness of the full inference pipeline under embedded execution with an inference latency of approximately 0.15 ms/sample and a memory footprint of only 0.40 MB. Altogether, the developed models offer an effective, interpretable, and computationally-efficient approach for PQD detection under evolving grid conditions, providing a standard of comparison for future approaches in PQD analysis.