Improved freezing rain forecast using machine learning

Abstract

Freezing rain is one of the most damaging weather phenomena in winter or early spring in many parts of the world, affecting traffic, power lines and agriculture. Thus, reliable and computationally efficient prediction of its occurrence is urgently needed in weather forecast operations. However, there are different thermodynamic processes that can lead to freezing rain, resulting in unsatisfactory forecasting performance of the state-of-the-art Numerical Weather Prediction (NWP) models. Here a data-driven forecasting method for freezing rain using machine learning technologies is proposed. Observations of weather phenomenon collected from 2515 national weather stations of China for winter of 2016 to 2019 and the corresponding atmospheric predictors derived from ERA5 reanalysis are used. The prediction function is constructed based on the classification and regression tree, and the predicting variables include temporal and vertical profiles of fundamental thermodynamic and kinematic parameters from 500hPa to 1000hPa, with a total dimension of 2304. The LightGBM (Light Gradient Boosting Machine) framework is adopted to train our prediction model and an algorithm-level approach of modifying the loss function is used to address the imbalance of classes to improve forecasting skill. Results show that the data-driven prediction model, namely DDFR (data driven forecast of freezing rain), out-performs the benchmark NWP, i.e., ECMWF IFS product. It’s improvements in terms of TS score range from 120% to 258% depending on different forecast leading times, which range from 0-12 hours. In addition, DDFR is applied in an operational NWP model of China. The problem of domain adaptation is tackled and transfer learning method is employed to adapt the original DDFR to this NWP model. The effectiveness of such adaptation has been demonstrated by its performance on both training and testing datasets.