Machine learning bias correction and downscaling of urban heatwave temperature predictions from kilometre to hectometre scale

Abstract

Abstract The urban heat island (UHI) effect exacerbates near‐surface air temperature (T) extremes in cities, with negative impacts for human health, building energy consumption and infrastructure. Using conventional weather models, it is both difficult and computationally expensive to simulate the complex processes controlling neighbourhood‐scale variation of T. We use machine learning (ML) to bias correct and downscale T predictions made by the Met Office operational regional forecast model (UKV) to 100 m horizontal grid length over London, UK. A set of ML models (random forest, XGBoost, multiplayer perceptron) are trained using citizen weather station observations and UKV variables from eight heatwaves, along with high‐resolution land cover data. The ML models improve the T mean absolute error (MAE) by up to 0.12°C (11%) relative to the UKV. They also improve the UHI diurnal and spatial representation, reducing the UHI profile MAE from 0.64°C (UKV) to 0.15°C. A multiple linear regression performs almost as well as the ML models in terms of T MAE, but cannot match the UHI bias correction performance of the ML models, only reducing the UHI profile MAE to 0.49°C. UKV latent heat flux is found to be the most important predictor of T bias. It is demonstrated that including more heatwaves and observation sites in training would reduce overfitting and improve ML model performance.

Keywords

• crowdsourced data
• land cover
• machine learning
• numerical weather prediction
• urban heat island