Long-lead term precipitation forecasting by Hierarchical Clustering-based Bayesian Structural Vector Autoregression

Kaixun Hua, Dan A. Simovici

Abstract: Heavy precipitation for several days and weeks always leads to some extreme nature disasters. Long-lead term precipitation forecasting plays an important role on the prevision of such calamities. Most works focus on the generation of training labels with allocation of the proper corresponding spatio-temporal information. In this paper, we will provide a different path by performing regression analysis using the precipitation amounts at particular locations. This method is called Hierarchical Clustering based Bayesian Structural Vector Autoregression (HC-BSVAR). The approach for HC-BSVAR is divided into two steps. First, we apply a hierarchical clustering algorithm to identify the Elite locations and then transfer the 3-dimensional data space into a new traditional 2-dimensional data space. Every column of the new data frame is a hydro-meteorological feature of the original data and each row represents a time point (day) in the original space. Secondly, an economic-based multivariate time series model called Bayesian-based Structural Vector Autoregression (BSVAR) is exploited to perform the final prediction result. The prediction quality will be vary for different cut of tree structure which generated by hierarchical clustering. The coefficient for determination of each location by each level of cut is applied to quantize the quality of prediction. The relationship between the cut level of clustering geographic locations and the regression model performance are also discussed, based on the result of prediction quality.