This work investigates the role of artificial intelligence (AI) in predictive maintenance for smart meters and photovoltaic (PV) systems through anomaly detection and forecasting. Three datasets were analyzed using unsupervised clustering, Random Forest classification, deep learning (LSTM), and statistical models (SARIMAX, Prophet). The solution achieved near-accurate forecasting (12-13% MAPE for short-term LSTM, 17-18% MAPE for PV energy) and nearly flawless anomaly categorization (ROC-AUC = 0.999; F1 up to 0.87-1.0). Anomalies including energy shortages, peak suppression, and inverter underperformance were translated into actionable maintenance insights through rule-based diagnostics and permutation-based feature importance analysis. These results demonstrate how AI- driven anomaly detection can significantly improve maintenance efficiency, reduce downtime, and enhance grid reliability by shifting from reactive repairs toward proactive, data-driven interventions.
Predictive Maintenance, Anomaly Detection, Smart Meter Data, Time-Series Forecasting, Graph Convolutional Autoencoder, LSTM, CNN–LSTM Autoencoder, Explainable AI, Hyperparameter Optimization, Residual Thresholding
© IACyC is conducted under the administration of the CyberMACS Erasmus-Mundus Masters Programme. cybermacs.org