A Hybrid Machine Learning Approach for Accurate Weather Forecasting in Dhaka City: Insights for Urban Planning and Disaster Management

Accurate weather forecasting is crucial for effective urban planning, disaster management, and public health decision-making, especially in rapidly urbanizing cities like Dhaka, Bangladesh. This study focuses on forecasting key weather parameters—temperature, humidity, and rainfall—for Dhaka City in 2026, using data from 2016 to 2025. We compare five advanced machine learning models—Random Forest (RF), XGBoost, CatBoost, Long Short-Term Memory (LSTM), and Time Series Transformer (TST)—to assess their forecasting accuracy and effectiveness. While Random Forest, XGBoost, and CatBoost demonstrated strong performance in identifying long-term weather trends, they struggled with capturing shortterm fluctuations. LSTM, a deep learning model, excelled in modelling temporal dependencies but was outperformed by TST. The TST model, leveraging an attention mechanism, showed the highest accuracy, capturing both long-term trends and short-term variations in the weather data. This ability to focus on relevant time steps makes TST particularly suitable for forecasting in dynamic, rapidly changing urban environments. The findings highlight the importance of adopting advanced machine learning models, such as TST, for improving weather prediction accuracy in Dhaka. Our research provides valuable insights for urban planners, disaster management authorities, and public health officials, offering a more reliable tool for managing the impacts of climate variability in Dhaka City.