A Comprehensive Framework for Predictive Maintenance in Solar Photovoltaic Systems Using Contextual Deep Model

This paper presents a framework for early fault prediction as part of predictive maintenance in solar photovoltaic (PV) systems using a context-sensitive deep learning approach. The proposed framework integrates a novel Context-Sensitive Long Short-Term Memory (CS-LSTM) regression model with a Random Forest-based classification module, leveraging both temporal and contextual features such as irradiance, temperature, time of day, and seasonal variations. Supervisory Control and Data Acquisition (SCADA) data from a 10 MW solar PV plant in India is used for model training and evaluation. The regression model estimates AC power under normal conditions, and deviations from actual measurements are flagged as potential faults. Fault classification is performed using a Random Forest model trained on imbalanced data balanced through Synthetic Minority Oversampling Technique (SMOTE), incorporating multiple fault types. Moreover, as a novel concept, an additional "Fault Impending" class has been introduced in order to provide early warning by the Random Forest classification module. Experimental results demonstrate that contextual integration significantly improves prediction accuracy, with the CS-LSTM model achieving an R²-score of 0.9741 and superior RMSE performance over baseline models. The Random Forest classifier attained the highest F1-score (0.992) in fault classification tasks. The framework also successfully predicted several faults up to 2 hours in advance, highlighting its potential for real-time monitoring.