Drivers and Predictive Modelling of Thunderstorms and Precipitation over West Africa

Abstract

Thunderstorms provide essential precipitation for agriculture and other economic activities in the tropics. Accurate forecasting of thunderstorm variability and precipitation forecasting over West Africa is vital for agricultural planning, water resource management, hydropower generation, and disaster risk reduction. This study aims to identify the trends and drivers of monthly thunderstorm frequency from 1990 to 2013 across 22 synoptic stations in Ghana using a machine learning approach. Additionally, for the first time, monthly thunderstorm frequencies over Ghana were forecast one year in advance using machine learning models to offer actionable insights for planning and disaster preparedness. Moreover, this study pioneers the application of machine learning to predict precipitation a year ahead over West Africa at a spatial resolution of 0.10° × 0.10°, filling a significant gap in long-term seasonal forecasting systems. LightGBM, XGBoost, Random Forest (RF), and ensemble models were developed utilizing ERA5 reanalysis data, climate indices, and geographic coordinates as input features. LightGBM and XGBoost models were selected for precipitation prediction due to their ability to efficiently process large datasets on standard computing infrastructure. To better understand model behaviour and the influence of features on monthly precipitation over West Africa, SHapley Additive exPlanation (SHAP) plots were employed. Recursive feature elimination was used to select the most important features from over 60 candidates to model thunderstorm variability, which the models explained at a rate of 79%. SHAP analysis revealed that convective available potential energy (CAPE) was the primary driver of thunderstorm variability, followed by top-of-atmosphere incident solar radiation (TISR), convective inhibition (CIN), 10 m wind speed (SI10), and longitude. The frequency of thunderstorms in Ghana declined significantly, coinciding with a notable decrease in CAPE and an increase in CIN, providing strong evidence of a causal relationship. The resulting LightGBM, XGBoost, RF, and ensemble models achieved R2 values of 0.74, 0.75, 0.75, and 0.76, respectively. This research introduces a forecasting scheme for thunderstorm frequencies in Ghana, offering a valuable tool for seasonal prediction. The precipitation models developed to forecast rainfall one year ahead across West Africa explained between 77% and 84% of the variance. Effectively capturing temporal and spatial rainfall patterns, including the ‘monsoon jump,’ ‘little dry season,’ and the Ghana-Benin dry zone, highlighting areas with high and low precipitation on a monthly basis. Analysis of model behaviour indicated that different features influence monthly precipitation across various zones of West Africa. These models can be adapted and integrated into the seasonal forecasts of national meteorological agencies to enhance accuracy. Better seasonal rainfall forecasts could assist West African farmers in anticipating rainfall variability, preparing for droughts and floods to safeguard crop yields. Countries within the study area that rely on hydroelectric power or surface water resources can also utilise these precipitation prediction models for planning and mitigating drought and flood risks.