基于机器学习的径流预测模型及适应变化的预测模式研究

Affected by changing environment, the evolution of runoff process becomes more complex, which always triggers controversy. So the research of runoff prediction which can rightly response and adapt to the environment changes is difficult but crucial. Based on the historical data and observations of actual changes, many problems are exposed in a traditional work,such as evaluation of consistency, reliability and representative. Driven by confronted problems, adopting feasible and advanced technical route is not only an exploration but also a breakthrough. In this study, the driving elements will be explored, the causes of changes will be revealed and the characteristic factors of runoff forecast will be extracted based on the big data analysis technology. Employed the causes as forward strategies and the results as reverse strategies, the runoff forecasting model will be established based on machine learning and deep learning which can response to environmental changes and improve the models applicability. The adaptive mechanism of runoff forecast will be built by compositing the big data predicting analysis, machine learning models and combination forecasting methods and integrated implement by knowledge mapping. Based on meta-synthesis platform and runoff forecast models component library, the runoff adaptive forecast system will be establised, the knowledge mapping of runoff forecast will be constructed according to combining different apply theme, forecast factors, time scales and models with each other. In this way, runoff forecast will be achieved by knowledge mapping. Through the study and try out, a runoff forecasting mode and integrated application mode will be proposed and formed, which will provide scientific support for the runoff prediction with changing environment.

受变化环境影响，径流过程的演变复杂，引发争议成常态，响应并适应变化环境的径流预测研究既难又关键。基于历史资料和实际变化调查，评价一致性、可靠性、代表性的传统工作暴露出了不少问题。由问题驱动，走可行又先进的技术路线是探索也是突破。基于对径流演变的大数据分析，挖掘驱动要素、揭示变化成因、抽取径流预测特征因子；采用机器学习和深度学习，以成因为正向策略和以结果为反向策略，建立能响应环境变化的径流预测模型，增强模型的适应性；综合大数据预测分析、机器学习模型和组合预测方法，建立应对各种变化的径流预测适应性机制并图谱化集成实现；基于综合集成平台及径流预测模型方法组件库，建立径流适应性预测系统，搭建不同应用主题、不同因子、不同时间尺度、不同模型相互组合的径流预测知识图，形成径流预测知识图谱，实现图谱化预测服务。通过研究及试用，形成适应变化的径流预测范式和集成应用模式，为变化环境下的径流预测提供科学支撑。

变化环境下，径流的演变过程愈发复杂，“稳态流域”这一假设不复存在，传统的水文工作方法与水文模型不能适应径流的非线性与变异性过程，且没有将预报真正用到实际工作中。本课题基于以上问题，重点开展了基于大数据挖掘的径流预测特征因子抽取，基于成因的径流预测机器学习模型，响应变化的径流适应性预测机制，径流适应性预测集成实现与实例应用等方面的研究。首先通过对于径流的成因分析，采用机器学习方法提取径流预报因子；针对泾河、渭河与汉江流域不同断面的水文数据，通过信号分解与机器学习算法的组合，构建了构建了多套机器学习径流预测模型及分解集成组合预报模型；对适应性、过程化与区间化等机制进行了研究，建立了响应变化的径流预测机制；利用知识图谱技术，对模型方法进行组件封装，集成实现了响应变化的径流预报平台。研究结果表明，基于成因的径流预报因子提取能良好的反映径流的变化规律；分解集成预报模型能更好的挖掘径流序列的隐藏信息，在不同的断面都取得了较好的效果；基于适应性机制的预报方法体系能更好的适应预报变化环境下的径流过程；基于适应性机制的径流预报业务集成应用，可真正将径流过程化动态预报的思想、方法和技术体系落到实处，能够将径流预报成果应用到实处。通过本研究，形成适应变化的径流预测范式和集成应用模式，可为变化环境下的径流预测提供科学支撑。