“…creek measured at St.Sulpice station (Ar) is used as a target variable, while the same water flow measured at Boudry station (Ar2) and the Seyon creek sharp fluctuations around the local trend due to instrumental errors. To remove this kind of artifact, the following preprocessing treatment is applied Oriani (2015): given a time-series Z(t) and computing the differential operator δZ(t) = Z(t) − Z(t − 1), the artifacts are identified with the portions of Z(t) presenting σ(t, a) > b, where σ(t, a) is the local standard deviation of 65 δZ(t), computed on the time interval [t ± a] and b is a user-defined threshold.…”