Comparison of Deep Learning Models in Predicting Water Deficits in Semi-Arid Regions: A Case Study of Dodoma, Tanzania

The escalating global freshwater shortage is driven by socio-economic development, changing consumption patterns, and systemic inefficiencies, with semi-arid regions like Dodoma, Tanzania, being especially vulnerable. Traditional statistical and regression-based models for predicting water deficits have proven insufficient in capturing the complex, nonlinear interactions among climatic, hydrological, and anthropogenic factors. To address this gap, this study proposes a deep learning-based predictive framework utilising advanced algorithms, Long Short-Term Memory (LSTM), Convolutional Neural Networks (CNN), Recurrent Neural Networks (RNN), and Deep Neural Networks (DNN) to improve the forecasting of water deficits. Using a thirteen-year (13) dataset collected from the semi-arid climate region of Dodoma, encompassing meteorological, hydrological, and socioeconomic variables. The models were trained and evaluated using performance metrics such as Root Mean Square Error (RMSE) and R-squared (R²). The DNN model demonstrated superior performance with an RMSE of 0.049 and an R² of 1.000, significantly outperforming other models. LSTM, CNN, and RNN models showed moderate to weak predictive accuracy, particularly in handling long-term dependencies and extreme deficit events. The key finding of this study is that the DNN model provides highly reliable and accurate water deficit predictions, making it the most effective among the tested deep learning approaches. This result highlights the value of incorporating deep learning into water resource planning, especially in data-scarce, semi-arid regions. The study concludes that DNN-based models should be prioritised for operational deployment in early warning systems and decision-making platforms. Future work should explore hybrid architectures, hyperparameter tuning, and integration with real-time data sources to enhance robustness and applicability