Short term storm intensity forecasting: a comparison of deep learning and machine learning methods

Climate risk assessment has become an increasing concern for insurance and reinsurance companies over the past decades. More specifically, the short term evolution of tropical storms intensity is of prime interest for local authorities to take preventive actions to protect populations at risk. In this paper, we present a deep learning model to forecast the intensity of tropical storm on a short term horizon, from 24 to 72 hours horizon. This is a challenging objective due to the complexity of time dependent and spatio-temporal features, such as the storm position, storm type (tropical, subtropical, extratropical, etc.), distance to coast, wind speed and meteorological data such as temperature, pressure, or sea surface temperature. The proposed deep learning forecasting model shows the potential of this approach for temporal analysis with recurrent neural networks (25% performance improvement compared to a basic and naïve model) and spatial features exploitation with convolutional neural networks, able to lear