Hourly Urban Air Temperature Forecasting with Graph Machine Learning

The Urban Heat Island (UHI) effect, where urban areas exhibit higher temperatures than their rural surroundings, is a growing subject of concern due to its implications for human health, energy demand, and anthropogenic emissions. Accurate high‐resolution forecasts of the UHI intensity and, more generally, of urban air temperatures are therefore crucial for guiding mitigation and adaptation strategies, especially for real-time heat warning systems and reliable power load forecasting. Here we present a novel spatio‐temporal graph machine learning framework for hourly urban air temperature forecasting that combines a Diffusion Convolutional Recurrent Neural Network (DCRNN) encoder with a multi‐horizon Multi-Layer Perceptron (MLP) decoder. The model is trained and tested using a dense network of 113 low‐cost temperature sensors in the city of Bern, Switzerland. A directed, weighted sensor graph is constructed using a composite distance metric of geographic proximity and environmental similarity, after which spatial dependencies are modeled using diffusion convolution and temporal dynamics via gated recurrence. The model is used to forecast the temperatures at each sensor with a horizon of 24 hours. Simulation results achieve an average root mean square error of 1.4K at 1 h and 2.44K at 24 h lead times, improving by up to 32% the baseline predictions of a per‐sensor Recurrent Neural Network (RNN). The practical application of the proposed methodology is further demonstrated by applying a regression‐kriging technique to generate hourly air temperature maps at 50 m resolution. Our results highlight the potential of graph‐based deep learning to deliver accurate, city‐scale, temperature forecasts that can be used as a high-resolution early warning system to enable interventions and informed decision-making for urban planning, public health, and emergency response in the case of extreme heatwave events.