A procedure for the detection of undocumented multiple abrupt changes in the mean value of daily temperature time series of a regional network

This paper presents the new procedure MAC-D for the automated detection of undocumented Multiple Abrupt Changes in the mean value of Daily temperature series, recorded in a network of meteorological stations. MAC-D can be applied to series containing seasonality, multiple change points, outliers, and with a noise component that can be autocorrelated and non-normally distributed. The main novelties of the procedure are (1) the pretreatment of the observed series, to derive a series of daily values that complies with the theoretical requirements of the change point detection tests and in (2) the combined use of the reference series and pairwise comparison approaches. MAC-D consists of three phases in sequence. In phase 1, the seasonal and climatic fluctuations are estimated and removed, using the reference series approach. Phase 2 combines a linear filtering with a change point detection test in an iterative algorithm, which runs until full compliance between the characteristics of the filtered series and the test requirements is achieved. Phase 3 is aimed at removing the false change points, due to error propagation in the reference series analysis, by double checking the detected change points with the pairwise comparison approach. Monte Carlo estimations of the actual significance and overall performance of the procedure for different series features and test resolutions are provided. Results demonstrate that MAC-D performs very well with daily series having a wide range of different characteristics.