Abstract:
Earth system consists of complex physics-based models to cover the various physical properties and stochastical patterns. In order to overcome the high complexity of existing methods, several studies attempt to predict the weather based on a large amount of observation data. However, they only focus on improving the prediction performance using deep learning models, and few studies analyze various meteorological characteristics. In this paper, we analyze what meteorological characteristics should be reflected in deep learning techniques to understand atmospheric phenomena. To predict weather conditions, we consider the following characteristics: (i) spatial features, (ii) temporal features, and (iii) the diversity of scale and shape of meteorological phenomena. The convolutional neural network is extended for extracting spatial patterns in the three-dimensional atmospheric fields, and a convolutional attention module is adopted to reflect the distinguished form of each atmospheric phenomenon. Then, dynamic patterns are extracted by the recurrent neural network-based time series model. We evaluated the performance of the proposed model with existing models by predicting the forecast fields. The experimental results showed that the synergy of various meteorological properties has a positive effect on weather prediction.

Keywords:
Weather Forecast · Meteorological Feature Analysis · Spatiotemporal Prediction · Convolutional Attention module

Citation:
Jeon, H.-J., Kang, J.-H., Kwon, I.-H. (2022).What Meteorological Characteristics Do Affect Weather Forecasts?,Computing4Human, 3, 7000.