Abstract
We compare modern machine learning architectures on the problem of long-term prediction of the air temperature at atmospheric pressure level 850hPa, denoted T850, for five locations of interest in the Norwegian Arctic sea. Our motivation comes from the utility of T850 for estimating long-term sea-spray icing risk, as developed in (Samuelsen et al. 2019, Section 5). Our comparison includes LSTM, N-HiTS, Temporal Fusion Transformer, and PatchTST. These architectures are compared against climatological normals and a (classical) statistical baseline consisting of MSTL and Prophet.
Our results show that some architectures which have shown promise for time-series weather forecasting fail to perform on the longer time horizons we consider. Other machine learning models are able to beat classical statistical methods, and we find it interesting that they appear to learn different features in the historical time-series in order to make their predictions. Our benchmark is carried out using both metrics from climatology and general statistics.
Our results show that some architectures which have shown promise for time-series weather forecasting fail to perform on the longer time horizons we consider. Other machine learning models are able to beat classical statistical methods, and we find it interesting that they appear to learn different features in the historical time-series in order to make their predictions. Our benchmark is carried out using both metrics from climatology and general statistics.