State of Health Forecasting of Lead-Acid Batteries Using Auto-regressive Integrated Moving Average (ARIMA) and Generalized Additive Models (GAM)

This thesis explores data-driven methods for estimating and forecasting the State of Health (SoH) of lead-acid batteries used in telecommunication systems. Traditional SoH assessment often relies on full discharge tests, which are time- and resource-intensive. Instead, this study uses operational performance data to identify variables associated with battery degradation and evaluate the predictive performance of three statistical models: Autoregressive Integrated Moving Average (ARIMA), ARIMAX (ARIMA with exogenous variable), and Generalized Additive Models (GAM) that incorporate the previous SoH value as an offset. The models were evaluated using rolling window cross-validation. ARIMA delivered the most accurate forecasts in terms of Root Mean Squared Error (RMSE) and Mean Absolute Error (MAE), outperforming both GAM and ARIMAX. The results suggest that simpler models may be more reliable when data are limited. With access to more data over a longer period, more advanced methods, such as machine learning, could improve SoH forecasting. This could also enable the prediction of alarm events, supporting more proactive and data-driven maintenance strategies in telecommunication systems.