ABSTRACT
A reliable urban extreme air temperature forecast in summer is necessary to effectively prepare for and mitigate the damage of thermal disasters such as heatstroke and power outages. A numerical weather prediction (NWP) model is commonly practical for operational forecasting, but it has forecast uncertainty due to coarse spatial resolution and unstable parameterization. This chapter introduces novel methodologies to improve urban air temperature forecasting performance through a fusion of an NWP model with machine learning. Two case studies conducted in Seoul, South Korea, are presented. They have different purposes: first is to produce spatial distribution maps of air temperature forecasts with improved accuracy; second is to evaluate the robustness of deep learning models for updating an operational NWP model. The first case study produced accuracy-improved air temperature maps through random forest, a support vector machine, and an artificial neural network. In addition, it improved the generalization ability using the multi-model ensemble that aggregates the outputs from those three machine learning models. The second case study demonstrated that a convolutional neural network that uses spatial information can be more robust to bias changes in operational NWP models. Based on the findings of these two case studies, this chapter highlights how machine learning can help further improve urban air temperature forecasting accuracy for operational use.
