Abstract Accurate prediction of extreme rainfall and dry events remains a major challenge in West Africa due to the complex and non‐linear dynamics of the monsoon system, which involve interactions among local convection, large‐scale circulation, ocean‐atmosphere coupling, and intra‐seasonal variations. Here, we develop a convolutional neural network (CNN) architecture that utilizes the vertically integrated moisture flux convergence (VIMFC) as predictor to discriminate between wet, normal and dry extreme conditions. Our model demonstrates high effectiveness in identifying wet (79%) and dry (82%) events, and captures the spatiotemporal variability. Interpretability diagnostics reveal that the CNN learns physically coherent patterns of VIMFC, consistent with established West African monsoon drivers. This data‐driven approach complements traditional forecasting and highlights the promising potential of deep learning for regional climate projections in data‐scarce regions. By linking statistical learning with physical understanding of the climate, our results suggest improved forecasting capability for climate risk management in West Africa.