EFFECT OF PENALTY FUNCTION USE ON CHANGE - POINT INVESTIGATION RESULTS FOR AIR TEMPERATURE DATA AT NHA BE AND CAN THO METEOROLOGICAL STATIONS

Abstract

This study aimed at the difference in change - point detection results for daily and weekly air temperature data using various penalty functions. Package "changepoint" in R was used to investigate the effect of penalty on the results of air temperature change - point assesment by mean for the period of 2013 ÷ 2017 at Nha Be and Can Tho meteorological stations. The results obtained by using PELT (Pruned Exact Linear Time) and SegNeigh (Segment Neighbourhood, for checking purposes) methods were identical. Meanwhile, the number of change - points found was increased gradually for SIC, HQIC and AIC penalty functions, respectively. The use of elbow - plot principle in combination with curve visual evaluation on graph demonstrated that the SIC function gives most reasonable number of change - points. It was also shown that the selection of time level (day or week) for input data will affect change - point investigation results. For weekly data, the change - points during the year were found to have very high coincidence (83.3 %, with the allowed deviation of 1 week) between Nha Be and Can Tho stations while that ratio was much lower if inputs belongs to daily type.