Optimizing Remaining Useful Life Prediction in Photovoltaic Systems Using AI-Enhanced LSTM Models

Blessed Oyebola; Emmanuel Babatunde

Abstract:

Introduction

Accurate prediction of the Remaining Useful Life (RUL) of photovoltaic (PV) systems is essential for proactive maintenance, cost reduction, and sustainable energy planning. Unlike structural health monitoring, RUL forecasting relies on modeling long-term temporal degradation patterns derived from electrical output, environmental exposure, and operational stressors. In Nigerian PV installations, high temperature variability, dust loading, and irradiance fluctuations intensify nonlinear degradation, rendering traditional statistical models insufficient. This work presents an AI-driven RUL prediction framework that systematically evaluates time-series learning models, with emphasis on Long Short-Term Memory (LSTM) networks integrated with Internet of Things (IoT) sensor data. The study is based on a systematic literature analysis and multi-criteria analytical evaluation of published photovoltaic diagnostic studies, rather than new experimental or field-generated data.

Methods

A multi-criteria analytical framework was applied to rank AI–hardware alternatives for RUL prediction using literature-derived, traceable performance metrics. Evaluated alternatives included RNN/LSTM + IoT sensors, ARIMA + sensors, ANN/DNN + I–V tracers, hybrid CNN + IoT + I–V ensembles, autoencoder-IoT models, and SVM-based approaches. Performance criteria comprised RMSE, MAE, MAPE, R², Accuracy, and Scalability. Metrics were normalized and weighted using the CRITIC method, which assigned dominant importance to R² (0.2632), RMSE (0.2584), MAE (0.2381), and MAPE (0.2367), while Accuracy received minimal weight (0.0036) due to limited discriminatory power. TOPSIS ranking was then applied to identify the most reliable RUL forecasting pipelines under real-world conditions.

Results

The TOPSIS evaluation ranked RNN/LSTM + IoT sensors as the strongest RUL prediction approach with a closeness coefficient of 0.663, significantly outperforming all alternatives. LSTM-IoT models achieved RMSE values as low as 0.0019–0.0020, MAE ≈ 1.24 kWh, and near-perfect explanatory power (R² ≈ 0.9999) across multiple studies. Hybrid CNN + IoT + I–V ensembles ranked second (Ci = 0.522), demonstrating robustness through multi-modal fusion. ARIMA-sensor models ranked third (Ci = 0.504), showing strong statistical fit but reduced adaptability to nonlinear degradation. Autoencoder-IoT methods reported extremely high accuracies (≈99–100%) but ranked lower due to weak error-based performance and limited field validation. ANN/DNN and SVM-based models consistently ranked lowest due to sparse reporting of RUL-specific error metrics and constrained scalability

Conclusion

The findings confirm that LSTM-based IoT frameworks are the most effective and field-ready solution for RUL prediction in photovoltaic systems, particularly under Nigerian environmental conditions. By minimizing prediction error while maintaining high explanatory power, LSTM-IoT models enable reliable forecasting of PV degradation trajectories and support predictive maintenance scheduling. When integrated with UAV-thermal SHM pipelines, the approach forms a unified, end-to-end diagnostic framework that addresses both immediate structural faults and long-term lifespan degradation, advancing scalable PV asset management in sub-Saharan Africa.