ARIMA-LSTM hybrid model for forecasting urban temperature dynamics in Ugandan cities

This study develops and evaluates an ARIMA-LSTM hybrid model for forecasting urban temperature dynamics in selected Ugandan cities, with the goal of capturing both linear trends and nonlinear fluctuations within a unified and interpretable framework. Monthly temperature data from 2017 to 2023 obtained from the Uganda National Meteorological Authority, alongside long-term urban population data from the World Bank, were used to support robust urban climate analysis. Data quality was verified through systematic preprocessing and outlier assessment, providing reliable inputs for model estimation. The proposed hybrid approach applies ARIMA to explicitly model-linear and seasonal temperature structures, while an LSTM network learns the remaining nonlinear patterns embedded in residuals. The model’s performance was evaluated against a wide range of benchmark models, including standalone statistical models, deep learning architectures, machine learning methods, and Facebook Prophet used strictly for comparison. The evaluation relied on multiple accuracy and goodness-of-fit measures such as RMSE, MAE, MAPE, SMAPE and R squared, complemented by visual diagnostics and classification-based performance analysis. Results consistently show that the ARIMA-LSTM hybrid outperforms all competing models, achieving smaller forecast errors, stronger explanatory power, better classification of temperature and more reliable prediction interval coverage. Forecasts generated for major Ugandan cities show persistent spatial differences in temperature patterns, with northern cities remaining warmer than highland regions. Overall, the findings demonstrate that hybrid modeling offers a reliable and practical tool for urban temperature forecasting, with clear relevance for urban climate planning, adaptation strategies, and evidence-based decision-making in Uganda.