基于机器学习方法对南京市月降水量建模预测（英文）

Modeling and prediction of monthly precipitation in Nanjing based on machine learning methods

摘要

【目的】准确的降水预测对区域防洪减灾、水资源管理及社会经济发展具有重要作用。然而，降水过程受多尺度气象因子相互作用影响，具有显著的非线性与时空异质性。传统数值模型难以有效捕捉其复杂演变规律。【方法】研究基于随机森林堆叠技术，构建了KNN-LSTM、SARIMA-KNN、SARIMA-Prophet、SARIMA-LSTM、Prophet-LSTM、Prophet-KNN六种组合预测模型，采用南京市58238站点1990—2023年月降水数据开展建模，其中1990—2020年数据被作为训练集，2021—2023年数据作为预测集。利用江苏省12个独立气象站点同期数据进行区域泛化能力验证。【结果】结果表明：SARIMA-LSTM组合模型通过融合SARIMA的季节性分解优势与LSTM的长期依赖特征捕捉能力，在预测集上表现出最优精度，其R²=0.904、MAE=16.16 mm和MSE=477.87 mm²。区域泛化能力验证显示该模型在江苏省13个气象站点的$ \overline{R^{2}}$=0.919、$ \overline{M A E}$=15.33 mm和$ \overline{M S E}$=537.52 mm²,表现出良好的空间泛化能力。【结论】构建的组合模型展现较好的预测性能，可为长江下游地区降水预测提供可靠的技术支撑，对区域水资源优化配置与灾害预警具有重要应用价值。

Abstract

Accurate precipitation prediction plays a crucial role in regional flood prevention and mitigation,water resources management,and socioeconomic development.However,the precipitation process is influenced by the interaction of multi-scale meteorological factors and shows significant nonlinearity and spatiotemporal heterogeneity.Traditional numerical models fail to effectively capture its complex evolution patterns. [Methods]Based on random forest stacking techniques,six hybrid prediction models were constructed:KNN-LSTM,SARIMA-KNN,SARIMA-Prophet,SARIMA-LSTM,Prophet-LSTM,and Prophet-KNN.Monthly precipitation data from 1990 to 2023 at station 58238 in Nanjing were used for modeling,with data from 1990 to 2020 used as the training set and data from 2021 to 2023 used as the testing set.The regional generalization ability was validated using contemporaneous data from 12 independent meteorological stations in Jiangsu Province. [Results]The result showed that the SARIMA-LSTM hybrid model,which integrated the seasonal decomposition advantage of SARIMA with the long-term dependency capturing ability of LSTM,achieved the highest prediction accuracy on the testing set,with R²=0.904,MAE=16.16 mm,and MSE=477.87 mm².The regional generalization validation demonstrated that the model achieved $ \overline{R^{2}}$=0.919,$ \overline{M A E}$ =15.33 mm,and $ \overline{M S E}$=537.52 mm² across 13 meteorological stations in Jiangsu Province,indicating good spatial generalization capability. [Conclusion] The constructed hybrid models exhibit excellent predictive performance,providing reliable technical support for precipitation prediction in the lower Yangtze River region.This holds significant application value for the optimization of regional water resource allocation and disaster early warning.

关键词

降水量预测 / 机器学习 / 组合模型 / SARIMA-LSTM / 区域泛化 / 月降水量 / 随机森林堆叠 / 影响因素

Key words

precipitation prediction / machine learning / hybrid models / SARIMA-LSTM / regional generalization / monthly precipitation / randomforeststacking / influencingfactors

基金资助

国家自然科学基金项目(52309025)