Precipitation nowcasting using ensemble learning approaches

The data from the automatic weather stations (AWS) though extremely important is yet to be fully explored from the perspective of weather forecasting. The proposed article experimented with a different setup of each ensemble learning technique XGBoost, AdaBoost and random forest with different oversampling techniques. The experiments lead us to develop an algorithm that is a linear combination of multilinear regression, XGBoost, AdaBoost and random forest. The predictors consist of time series of in-situ observations. We have also studied the impact of in-situ observations on the rainfall for the next few hours based on misclassification error. The results indicate that the most influential feature extracted from the proposed algorithm is humidity and rainfall while other meteorological variables are found to be weak predictors. The average accuracy of the proposed algorithm is 87%.