Data-Driven Deep Learning Emulators for Geophysical Forecasting

We perform a comparative study of different supervised machine learning time-series methods for short-term and long-term temperature forecasts on a real world dataset for the daily maximum temperature over North America given by DayMET. DayMET showcases a stochastic and high-dimensional spatio-temporal structure and is available at exceptionally fine resolution (a 1 km grid). We apply projection-based reduced order modeling to compress this high dimensional data, while preserving its spatio-temporal structure. We use variants of time-series specific neural network models on this reduced representation to perform multi-step weather predictions. We also use a Gaussian-process based error correction model to improve the forecasts from the neural network models. From our study, we learn that the recurrent neural network based techniques can accurately perform both short-term as well as long-term forecasts, with minimal computational cost as compared to the convolution based techniques. We see that the simple kernel based Gaussian-processes can also predict the neural network model errors, which can then be used to improve the long term forecasts.