AI-Enhanced Battery Degradation Prediction for Solar Home Systems in Sub-Saharan Deployment Conditions

Abstract

The durability and reliability of batteries in solar home systems (SHS) are critical to the long-term success of off-grid electrification efforts in Sub-Saharan Africa. However, harsh environmental conditions, variable load profiles, and limited maintenance capacity contribute to accelerated battery degradation and unexpected failures. This study presents a data-driven framework for accurate prediction of battery state of health (SOH) using advanced machine learning models under deployment-relevant conditions. Three architectures—Long Short-Term Memory (LSTM), eXtreme Gradient Boosting (XGBoost), and Transformer—were trained and evaluated using a high-resolution synthetic dataset simulating 1,000 battery cycles. The dataset incorporated temperature variability, depth of discharge (DOD), charge rate fluctuations, and measurement noise to reflect real-world SHS operating environments. Model performance was assessed using MAE, RMSE, and R2R^2R2 metrics. The Transformer model consistently outperformed others, achieving the highest accuracy and lowest error variance, with residuals tightly centered around zero. SHAP analysis revealed temperature as the dominant contributor to degradation, followed by DOD and charge rate. The deployment feasibility of each model was also validated through inference benchmarking on a Raspberry Pi 4, confirming sub-300 ms runtime and minimal memory consumption suitable for low-power edge computing. These findings establish the Transformer model as a viable candidate for real-time, embedded battery diagnostics in SHS applications. The integration of such AI-based prediction systems offers a scalable solution to enhance battery longevity, reduce maintenance costs, and ensure uninterrupted energy access in underserved regions. The approach also supports adaptive energy management, warranty validation, and sustainable SHS design.