Regional Forecast of Heavy Precipitation and Interpretability Based on TD-VAE

Heavy precipitation is a kind of regional weather condition with transitory action time, strong outburst and serious disaster. The formation of heavy precipitation weather generally requires: unstable thermodynamic and dynamic parameters and abundant precipitable water. This research solved the dilemma of forecasting severe precipitation in Tianjin. Firstly, the missing values in the original encrypted data are estimated according to the spatio-temporal interpolation method. Secondly, some convective parameters commonly used in meteorological field are calculated to increase the credibility of the classifier. Moreover, a variational autoencoder-based transfer discrimination (TD-VAE) algorithm is employed to tackle the difficulty of class imbalance: the TD-VAE generates a mass of different cases from the minority classes of the imbalanced dataset to train the classifier. Finally, the rationality of the calculated convection parameters and TD-VAE is demonstrated by an explainable machine learning method. The experimental results reveal the data augmentation strategy based on VAE can well deal with the dilemma of class imbalance in binary classification, which is superior to the traditional oversampling algorithm in several machine learning methods.