A Hybrid AI Decision Framework for Predictive Maintenance of Photovoltaic Systems with Formalized Human-in-the-Loop Co-Learning

ECH-CHARQAOUY, Amjad; ECH-CHARQAOUY, Sidi Salah; BOULEZHAR, Abdelkader; MIHRAMANE, Redouane; ECH-CHARQAOUY, Nizar

doi:10.22059/jser.2026.413937.1738

A Hybrid AI Decision Framework for Predictive Maintenance of Photovoltaic Systems with Formalized Human-in-the-Loop Co-Learning

Articles in Press

Document Type : Research Article

Authors

Amjad ECH-CHARQAOUY

Sidi Salah ECH-CHARQAOUY

Abdelkader BOULEZHAR

Redouane MIHRAMANE

Nizar ECH-CHARQAOUY

Department of Physical Sciences, Faculty of Sciences Ain Chock (FSAC), Hassan II University of Casablanca, Morocco

10.22059/jser.2026.413937.1738

Abstract

Predictive maintenance of utility-scale photovoltaic (PV) systems operating in harsh desert environments requires decision frameworks that simultaneously handle operational uncertainty, multi-criteria trade-offs, and limited human supervision. This paper proposes a hybrid Multi-Criteria Decision Support System (MCDSS) formulating PV maintenance as a sequential decision-making problem under uncertainty. The framework integrates fuzzy logic inference, an Adaptive Decision Tree (ADT), a Q-learning reinforcement module, and a human-in-the-loop (HITL) co-learning mechanism governed by three operator feedback types: Confirm, Modify, and Override. Validated on semi-real operational data from the Noor Boujdour I solar plant (Morocco), the system achieved an average reaction time below 8 s, a corrected availability rate of 97.3%, and a post-adaptation decision adequacy rate of 92.3% (N = 142; Cohen’s κ = 0.83, 95% CI [0.76, 0.90]). McNemar tests showed significant gains over rule-based, Random Forest, and ML baselines (p < 0.05), but not over standalone Q-learning (p = 0.343). Ablation analysis identifies HITL co-learning as the largest single contributor to performance gain (−4.4 pp without HITL, p = 0.012, McNemar). The proposed architecture demonstrates promising adaptability for renewable-energy maintenance applications, subject to further multi-season validation.

Keywords

Predictive maintenance

Human-in-the-loop learning

Multi-criteria decision support

Reinforcement learning

Photovoltaic systems