Urban Temperature Dynamics in Sri Lanka: A Comparative Machine Learning and Time-Series Analysis Across Wet, Intermediate, and Dry Climate Zones

Why It Matters

Accurate urban temperature forecasts empower agriculture, energy management and public‑health planning in rapidly urbanizing tropical regions, where climate variability is intensifying.

Key Takeaways

• SVR with RBF kernel outperformed all models, R² >0.98
• ARIMA with Fourier terms lagged behind machine learning techniques
• Temperature mean, max, min, and evapotranspiration drive forecasts
• Study covered 13 years of daily data across three climate zones
• Forecast errors as low as 0.11°C MAE, enabling precise planning

Pulse Analysis

Urban heat islands are a growing concern for tropical megacities, where rapid expansion and land‑use changes amplify temperature swings. Reliable daily temperature forecasts are essential for sectors such as agriculture, which depends on precise planting windows, and energy utilities that must balance cooling demand. Traditional statistical approaches often struggle with the non‑linear interactions inherent in these environments, prompting researchers to explore more adaptable machine‑learning solutions.

The Sri Lankan study leveraged a rich dataset of 4,916 daily observations per station, spanning 2010‑2023, across Colombo, Kurunegala and Puttalam. After rigorous outlier handling and rainfall transformation, the team trained multiple models, finding that support vector regression with a radial basis function kernel consistently outperformed both classical ARIMA‑Fourier and other ensemble methods. With R² exceeding 0.98 and MAE as low as 0.11 °C, the SVR model captured subtle temperature dynamics, while feature‑importance analysis identified apparent temperature, daily extremes and evapotranspiration as the dominant drivers.

These findings have immediate practical implications. Municipal planners can integrate the high‑precision forecasts into heat‑stress mitigation strategies, while power grids can fine‑tune load‑balancing algorithms to reduce outages during peak heat. Moreover, the demonstrated superiority of machine‑learning techniques over conventional time‑series models encourages broader adoption across other tropical urban contexts. Future research may expand the feature set to include satellite‑derived land‑surface metrics, further sharpening predictive power and supporting climate‑resilient urban development.