Abrupt change point detection of annual maximum precipitation using fused lasso

- This study proposed the fused lasso penalty function to detect the change point of AMP, which can generally be fitted using the GEV distribution.
- Four numerical experiments and the real data application for 63 weather stations are conducted to compare the detection performance between BCPA and the fused lasso method.
- The GEV fused lasso method detects the change points of the non-stationary AMPs of South Korea because the means of AMPs didn't go back to their previous values after the change point.

SummaryBecause the widely used Bayesian change point analysis (BCPA) is generally applied to the normal distribution, it cannot be freely used to the annual maximum precipitations (AMP) in South Korea. Therefore, this study proposed the fused lasso penalty function to detect the change point of AMP which can be generally fitted by using the Generalized Extreme Value (GEV) distribution in South Korea. First, four numerical experiments are conducted to compare the detection performances between BCPA and fused lasso method. As a result, fused lasso shows the superiority of the data generated by GEV distribution having skewness. The fused lasso method is applied to 63 weather stations in South Korea and then 17 stations having any change points from BCPA and the GEV fused lasso are analyzed. Similar to the numerical analyses, the GEV fused lasso method can delicately detect the change point of AMPs. After the change point, the means of AMPs did not go back to the previous. Alternately, BCPA can be stated to find variation points not change points because the means returned to their original values as time progressed. Therefore, it can be concluded that the GEV fused lasso method detects the change points of non-stationary AMPs of South Korea. This study can be extended to more extreme distributions for various meteorological variables.