Abstract: Drought is one of the major natural disasters,whose occurrence is linked to a sustained lack of precipitation.The drought forecast provides vital evidence and support for preventing losses of drought disasters,and therefore it is of great significance.In this study,a series of the standard precipitation evapotranspiration index (SPEI) at different time scales were calculated based on the daily precipitation and temperature data from 7 meteorological stations in Sanjiang Plain,northeast China from 1960 to 2016 and were used to forecast the drought using ARIMA and ANN models.In the stage of training and testing,the fitting degrees of the models were evaluated and validated and the optimal ARIMA and ANN models were chosen with the help of 6 fitting evaluation methods:the correlation coefficient (R),Nash-Sutcliffe efficiency coefficient (NSE),Kendall,rank correlation coefficient,the mean square error (MSE) and Kolmogorov-Smirnov (K-S) test.Then 12 values for the 12 months in 2016 were predicted by the optimal models and were compared with the corresponding observations.The results are shown as follows:(1) The prediction ability of ARIMA and ANN models based on SPEI were both increased with the increase of time scale in Sanjiang Plain.(2) The two models had poor prediction accuracy for SPEI 3 and SPEI 6.For the SPEI value of 9,12 and 24 months,all models worked well with accuracy more than 70%.(3) For the SPEI value of 9,12 and 24 months,the prediction accuracy of ANN model is better than that of ARIMA model.In particular,the prediction accuracy for one month forecast of SPEI 12 and 24 at all stations were more than 80%.All these showed that the prediction model of ANN has strong maneuverability and can effectively predict the drought at a large time scale in Sanjiang Plain.The drought prediction at small time scale (3 and 6 months) needs to be improved in future studies.

Key words: drought, ARIMA model, ANN model, SPEIs, Sanjiang Plain