降水量预测是现代气候预测业务的核心和难点。耦合模型在新疆降水量预测的研究应用屈指可数,因此,通过尝试建立集合经验模态分解（Ensemble Empirical Mode Decomposition,EEMD）和长短期记忆神经网络（Long Short-Term Memory Network,LSTM）的耦合模型对天山北坡经济带降水量进行预测研究。将1965—2019年天山北坡经济带共55 a的年降水量数据进行EEMD分解,转换成4个平稳分量和趋势项,通过谱分析得出各个分量的准周期,为后续训练LSTM模型提供基础。根据EEMD分解后的各分量训练得出LSTM网络模型并利用该模型进行研究区降水量预测。结果表明：EEMD-LSTM耦合模型预测2010—2019年天山北坡经济带降水量的平均相对误差为13.38%,均方根误差为38.03 mm,认为EEMD-LSTM耦合模型对天山北坡经济带降水量预测精度较好。利用EEMD-LSTM耦合模型预测2020—2029年天山北坡经济带年降水量,其中有6 a降水偏多,4 a降水偏少,2025年可能为极端湿润年,降水偏多超过20%;而2021年为极端干旱年,降水量预计低于200 mm。本文探索了干旱区降水量预测的新方法,并为气象防灾减灾工作提供参考依据。

Precipitation prediction is both an essential and challenging component of modern climate prediction. In the precipitation prediction in Xinjiang, the research and application of coupling models are very limited. Therefore, this paper attempts to establish the ensemble empirical mode decomposition (EEMD) and long short-term memory (LSTM) neural network coupling models to predict precipitation in the Northern Slope Economic Belt of Tianshan Mountains. Firstly, the precipitation data recorded in the study area during 55 years, from 1965 to 2019, were decomposed into four stationary components and trend terms using the EEMD, and the quasi period of each component was obtained by spectral analysis, which provided the basis for the subsequent training of the LSTM model. Then, each EEMD component was trained into the LSTM network model, and the models were used for predictions. The reconstruction and comparison of the results revealed that the average relative error and root mean square error of the 2010-2019 model were 13.38% and 38.03 mm, respectively. Therefore, it is inferred that the EEMD-LSTM coupling model can achieve a better precipitation prediction accuracy in the study area. The model was used to predict annual precipitation in the Northern Slope Economic Belt of Tianshan Mountains from 2020 to 2029; within this period, six years presented more and four years presented less precipitation. Year 2025 is predicted to be an extremely humid year with more than 20% of precipitation in excess; while 2021 is anticipated as an extremely dry year, with expected precipitation amounting to less than 200 mm. This study explored a new method of precipitation prediction in an arid area, and provided a reference for meteorological disaster prevention and mitigation.