Application of Generative Adversarial Networks (GANs) for Precipitation Forecasting

Abstract:

As extreme weather becomes more frequent due to increased variability in a changing climate, better prediction of extreme weather and water events is critical for water resources and emergency management. This study presents a collaborative operational process that leverages advanced Machine Learning techniques to improve physically-based earth system modeling using observations. We propose a methodology to forecast precipitation using conditional Generative Adversarial Networks (cGANs) which can account for the temporally evolving 3-dimensional hydrometeor field. Leveraging a 34-year reforecast data set (West-WRF) developed at the Center for Western Weather and Water Extremes (CW3E), an unprecedented opportunity is provided to train powerful artificial neural networks on historical data for the prediction of precipitation. Leveraging information from past model error over a wide range of atmospheric and hydrologic conditions and multiple vertical levels, cGAN improves 24-hour precipitation accumulation forecasts accuracy for up to 5 days leading time. The PRISM precipitation climate dataset is used as the ground truth to train the neural network and verification metrics such as Probability of Detection, False Alarm Ratio, Critical Success Index, correlation coefficient, and Root‐Mean‐Square Error are used to investigate the superiority of the proposed method compared to current operational precipitation prediction approaches over the states of California and Nevada.