Background Drought is a temporary and recurring meteorological event that has the most serious impact on human society. The arid region of Northwest China is located in the hinterland of the Eurasian continent. Due to severe insufficient precipitation and high evaporation, the arid region of Northwest China often suffers from the impact of drought, seriously affecting local agricultural production and life, and causing serious social and economic impacts. Establishing a drought prediction model applicable to the arid region of Northwest China will effectively reduce the impact of varying degrees of drought on these areas.

Methods In order to explore the neural network model suitable for meteorological drought prediction in the arid region of Northwest China, based on the precipitation data of twelve meteorological stations in the arid region of Northwest China from 1987 to 2016, the standardized precipitation index (SPI) was used as an indicator. According to the different input variables, nine groups of models were established based on the back-propagation neural network (BPNN), the extreme learning machine (ELM), and the long short-term memory network (LSTM) to predict the meteorological drought. And the stability of the model through the GLDAS dataset was verified.

Results 1) The prediction accuracy of ELM is slightly improved compared to BPNN, and the training time is shorter. However, ELM and BPNN have low prediction accuracy in some regions, low reliability of the model, and poor applicability. And the two models are difficult to maintain good prediction accuracy in the case of a single input variable.2) The analysis of the results of the meteorological drought prediction model show that the prediction accuracy of LSTM is higher than that of BPNN and ELM. The coefficient of determination (R²) of the highest accuracy model M7 is 0.965, and the root mean square error(RMSE) is 0.175. The R² in typical years are all greater than 0.8, and the RMSE is less than 0.132. 3)The analysis of the prediction results of typical years shows that LSTM performs well in the prediction of different typical years, and the prediction accuracy of dry year and extremely dry year is slightly higher than that of wet year and normal year. Meanwhile, validation of the GLDAS (global land data assimilation system) dataset showed that the best performing model M7 maintained an R² of 0.9 or above, indicating that LSTM can ensure stable and high-precision prediction of drought conditions in the arid region of Northwest China.

Conclusions Compared with BPNN and ELM, LSTM has stronger applicability and can still maintain higher prediction accuracy with fewer input variables. It shows stable predictive ability when facing different datasets. The above results indicate that the prediction model using LSTM has good applicability in the meteorological drought prediction of the arid region of Northwest China.