A comparative study of deep learning-based time-series forecasting techniques for fine-scale urban extreme heat prediction using Internet of Things observations

Human exposure to extreme heat has increased tremendously as climate change worsens. Extreme summer temperatures are breaking records around the United States, with multiple cities experiencing their hottest summer months in history. Extreme weather events such as wildfires in the western US only exaggerated the extreme heat with power outages on sections of the electrical grid. As the climate continues to warm, urban areas with growing populations will suffer more frequently and intensely from these situations. To better assist urban citizens in their daily activities, localized, street-level weather forecasts are needed because air temperature can have higher variability than the regional forecast. Situations are especially challenging to predict when the weather changes rapidly. In this research, air temperature observations were collected at a very fine scale (~150m) along major roads of New York City (NYC) through the Internet of Things (IoT) data provider GeoTab for 2019-20. To better learn the temporal patterns in air temperature within NYC over the past year, we used historical weather observations from in situ weather sensors (i.e., Weather Underground, WU) within the region for the past 5 years. We compared three sequence-to-sequence models based on deep neural networks, i.e., stacked long short term memory (LSTM) model, convolutional gated recurrent networks (ConvGRU), and Transformer to predict air temperature for the future 24h based on the past 48-h observations, either from GeoTab only or GeoTab+WU. Experiments of models trained using GeoTab observations showed that the ConvGRU performed the best within all considered models and integrating historical weather observations did not necessarily improve the predictive performance.