Abstract: Abnormal climate events have demonstrated an increasing trend with global warming in recent years. Continuous abnormal climate events refer to the phenomenon that weather/climate state constantly deviates from the average status. Compared with the traditional definition of abnormal events, continuity and overrun of continuous abnormal climate events have been often overlooked, but they also seriously affect the production and life of the society. Aiming at filling the gap that traditional anomaly monitoring methods can not detect continuous abnormal weather, this paper firstly presents a probability-percentile algorithm that adopts the continuous abnormal monitoring idea with continuous large deviation from suitable value. On this foundation, gated recurrent unit (GRU) neural network was applied to predict continuous abnormal meteorological value. The model was applied to daily meteorological data of precipitation, temperature, and wind speed at 166 stations in mainland China from 1951 to 2020, and the results suggest that as the duration increases, continuous abnormal meteorological value presents a fluctuating pattern in most regions, rather than a hypothetical downward trend. Therefore, significant attention should be paid to continuous abnormal weather with high duration and average daily meteorological value based on our model. The method proposed in this paper can be used to monitor and predict continuous abnormal climate events, and is a valuable supplement to traditional anomaly detection methods.

Key words: continuous anomaly, probability-percentile algorithm, anomaly detection, GRU, meteorology