Deep Prediction Interval for Weather Forecasting
share
Currently there exists a gap between deep learning and the techniques required to quantify uncertainty. This paper uses the weather forecast as an application background to illustrate the technique of deep uncertainty learning (DUL). Weather forecasting has great significance throughout human history and is traditionally approached through numerical weather prediction (NWP) in which the atmosphere is modeled as differential equations. However, due to the instability of these differential equations in the presence of uncertainties, weather forecasting through numerical simulations may not be reliable. This paper explores weather forecasting as a data mining problem. We build a deep prediction interval (DPI) model based on sequence-to-sequence (seq2seq) that predicts spatio-temporal patterns of meteorological variables in the future 37 hours, which incorporates the informative knowledge of NWP. A big contribution and surprising finding in the training process of DPI is that training by mean variance loss instead of mean square error loss an significantly improve the generalization of point estimation, which has never been reported in previous research. We think this a new kind of regularization can not only be on a par with the famous Dropout but also provide more uncertainty information, which comes into a double-win situation and we name it MV-R (Mean Variance Regularization). Based on single DPI, we then build a deep ensemble. We evaluate our method on 10 realistic weather stations in Beijing of China. Experimental results shown DPI has better generalization than traditional point estimation and deep ensemble can further improve the performance. Our deep ensemble method achieved top-2 online score ranking in the competition of AI Challenger 2018. It can dramatically decrease up to 56 NWP, which has important use value and practical significance.
READ FULL TEXT
