Short-term rainfall forecast by several typical machine learning algorithm

According to the characteristic changes of precipitable water vapor and meteorological parameters (temperature (T), humidity (U), dew point temperature (T_d), surface pressure (P)) during the rainfall process, it is possible to establish a short-term rainfall forecast model based on machine learning algorithms. This paper uses the 3-hour zenith tropospheric delay and meteorological data of the bjfs station and wuh2 station in 2020 as examples to construct the prediction model of the four algorithms: random forest (RF), support vector machine (SVM), K-nearest neighbor (KNN), and naive bayes classifier (NBC), and introduces the rainfall events at each time as the new feature vector, adopts the segmentation method of 70% and 80% training sets respectively, takes the rainfall events as the model output, and the applicability of the model is evaluated by the accuracy, precision rate and false negative rate. After obtaining the accuracy is about 0.92, the precision rate is about 80%, and the false negative rate is about 20%, the data of 150—200 days in the time series are further used as samples to predict the rainfall of 200—250 days. The results indicate that The short-term rainfall forecast model based on machine learning can predict more than 80% of the rainfall events in the next 3 hours, and the false negative rate is below 20%, among which the SVM model has better comprehensive performance. Compared with the traditional threshold model, the accuracy rate is equivalent, and the false negative rate is decreased by about 50%.