Comparison of the Ability of ARIMA, WNN and SVM Models for Drought Forecasting in the Sanjiang Plain, China

Abstract

Drought is a natural disaster that profoundly impacts all parts of the environment. Drought forecasting could provide technical support for drought risk prevention. This paper explored and compared the forecasting abilities of the autoregressive integrated moving average (ARIMA) in statistics, the wavelet neural network (WNN) and the support vector machine (SVM) in machine learning for drought forecasting in the Sanjiang Plain, China. The models used in this paper are based on the standard precipitation evapotranspiration index (SPEI) on the 12-month timescale. The SPEI was calculated using precipitation and temperature data collected during the period 1979–2016 from seven meteorological stations in the study area. Then, the SPEI series were predicted with the ARIMA, WNN and SVM models separately. The coefficient of determination (R2), mean-squared error (MSE), Nash–Sutcliffe efficiency coefficient (NSE) and Kolmogorov–Smirnov (K–S) distance, which is a nonparametric measure, were used to evaluate the performance of all models. A comparison between the raw data and predictions illustrates that the R2 and NSE values of the WNN model were 0.837 and 0.831, respectively; those of the SVM model were 0.833 and 0.827, respectively; and those of the ARIMA model were both > 0.9. Moreover, the ARIMA model had smaller MSE and K–S distance values than those of the other two models. Further, analysis of variance showed that the ARIMA model exhibited an obvious advantage over the other two models when forecasting drought in the Sanjiang Plain, China. Therefore, the method used for drought forecasting depends not only on the object of the data series but also on the underlying concepts of the models or algorithms and is a choice that should be made with caution.