Using model based AI to improve operational predictability of heat waves in developing countries: case study of Abidjan

Abstract

Heat waves (HW) pose significant threats as deadly natural disasters, leading to human casualties and crop damage. The warming trend attributed to climate change is particularly affecting tropical Africa, with further temperature increases expected in the future. Consequently, HW events are projected to become more frequent in this region. This study first provides a temporal investigation of HW occurrences in Abidjan from 2009 to 2022, shedding light on its trend during these past years. In this work, an approach is proposed to predict HWs over Abidjan. The approach is based on a deep learning (DL) model that is optimized through Bayesian optimization and trained with the fifth generation of European ReAnalysis (ERA5) and historical data from the Abidjan synoptic station. Furthermore, because of the scarcity of this event caused by its unusual occurrence, the proposed approach leverages the advantages of transfer learning (TL) and random under-sampling (RUS) techniques to effectively address the challenge of class imbalance in the available data. The model demonstrates a remarkable performance which is supported by an AUC metric value of 99.4% for a RUS rate of 0.25% indicating high discriminatory power and predictive accuracy. This study gives important insights into HW prediction in West African capitals and highlights the effectiveness of artificial neural networks (ANNs) for effective mitigation of socio-economic impacts.