A Comparative Approach to Batch and Online Machine Learning Methods for Real-Time Weather Forecasting: The Case of Cameroon

This study investigates the performance of multiple machine learning models, operating under both Batch and Online learning paradigms, for temperature forecasting across six climatically representative cities in Cameroon: Douala, Ebolowa, Bafoussam, Ngaoundéré, Garoua, and Maroua. The models evaluated encompass the Batch Linear Regressor, Gradient Boosting, Random Forest Regressor, Hoeffding Tree Regressor, Adaptive Random Forest Regressor, and Online KNN as well as their Online counterparts. The experimental results demonstrate that forecast accuracy, as measured by the Mean Absolute Error (MAE), is strongly contingent upon the prevailing climatic context. A discernible South-to-North performance gradient emerges: prediction errors increase progressively with the intensity of seasonal variability and thermal extremes, spanning from Douala (humid equatorial climate) to Maroua (hot Sahelian climate). The study establishes the consistent superiority of Online models over their Batch counterparts, attributable to their inherent capacity to adapt continuously to climatic fluctuations and seasonal concept drift. Among all models evaluated, the Adaptive Random Forest Regressor achieves the highest predictive accuracy, while the Online Linear Regressor offers the most favorable overall trade-off between accuracy, computational speed, and memory efficiency rendering it particularly well-suited for large-scale operational deployment. This research reaffirms the imperative of aligning predictive models with local climatic characteristics, and demonstrates that incremental learning constitutes a robust and adaptive framework for temperature forecasting in environments subject to pronounced seasonal variability, such as those encountered across Cameroon.