Improving BP artificial neural network model to predict the SPI in arid regions: a case study in Northern Shaanxi, China

Abstract

Drought prediction plays an important guiding role in drought risk management. The standardized precipitation index (SPI) is a popular meteorological drought indicator to measure the degree of drought. The SPI time series is non-stationary, whereas the conventional artificial neural network (ANN) model has limitations to predict non-stationary time series. To overcome this limitation, it is essential to investigate input data preprocessing to improve the ANN model. In this paper, a hybrid model coupled with singular spectrum analysis (SSA) and backpropagation ANN is proposed (SSA-BP-ANN). The advantage of this model is that the SSA of finite-length SPI sequences does not require the adoption of boundary extensions to suppress boundary effects, while the most predictable components of the SPI can be efficiently extracted and incorporated into the model. The proposed SSA-BP-ANN model is tested in case studies at three meteorological stations in Northern Shannxi Province, China. The results show that the SSA-BP-ANN model can produce more accurate predictions than the BP-ANN model. In addition, the performance improvement of SSA on the BP-ANN model is slightly better than wavelet decomposition and empirical mode decomposition. This new hybrid prediction model has great potential for promoting drought early warning in arid regions.