基于知识学习的洪水预报方法研究

Current flood forecasting mainly adopt the combination way of flood forecasting model and forecasting experience. However, forecasting experience mostly depends on the forecaster’s experience, it is difficult to build up knowledge base considering all kinds of change conditions in order to realize intelligent flood forecasting. Introducing the knowledge learning method in the filed of flood forecasting and studying the methods of forecasting experience learning and forecasting knowledge accumulation from flood are of great significance for improving flood forecasting theory and technology. Through the procedures below, including choose of typical basin, feature recognition of basin history flood, choose of corresponding typical flood, mining of rainfall runoff variation characteristics under the change condition, and then the description of forecasting knowledge from rainfall runoff variation characteristics and generation mechanism of prediction error, this project studies forecasting knowledge iterative learning method based on historical flood classification process, also establishes flood forecasting knowledge base under the change condition. Finally a systematic and theoretical framework is developed for flood forecasting method base on knowledge learning, i.e., physical mechanism analysis, forecasting knowledge description, forecasting knowledge learning, forecasting knowledge base building, forecasting scheme formulation and revision, which provides a scientific basis for effectively utilization of flood resources.

目前洪水预报多采用洪水预报模型和预报经验联合使用的方式，预报经验较多依赖于预报员的自身经验与水平，难以建立考虑各种变化条件的洪水预报知识库以实现变化条件下的洪水预报智能化。因此将知识学习引入到洪水预报领域，研究从各次或各类洪水中学习预报经验与积累预报知识的方法，对完善洪水预报理论与技术具有重要意义。本项目拟通过典型流域选取、流域历史洪水一般特征识别以及相应典型洪水选取，挖掘变化条件下的降雨径流演变变异特征，通过基于降雨径流演变特征及预报误差产生机制的预报知识描述，研究基于历史分类洪水过程的预报知识迭代学习方法，建立变化条件下的洪水预报知识库，最终形成一个基于知识学习的洪水预报方法的系统理论框架，即“物理机制分析—预报知识描述—预报知识学习—预报知识库建立-预报方案制定与修正”，为更有效地进行洪水资源利用提供科学依据。

本项目首先基于耦合多源降雨数据的流域洪水预报模型，分析降雨径流演变特征及预报误差，建立了新安江-BP神经网络耦合模型；然后，基于流域WRF中尺度数值天气模式与VIC大尺度分布式陆面模型的建立，挖掘出流域分布式洪水预报的影响因素，实现了WRF中尺度数值天气模式与VIC大尺度分布式陆面模型的耦合；利用多源遥感信息定量分析上游水利工程水库在流域场次洪水中的拦蓄影响及拦蓄规律，实现考虑上游水利工程影响的流域洪水预报；在此基础上，建立的基于长短时记忆网络（LSTM）和深度学习堆叠架构的流域洪水预报模型，充分利用了洪水演变特征与预报误差产生机制的预报知识，迭代学习不同预见期下降雨特征因子、径流特征因子以及二者交互作用在洪水过程中对输出流量的作用，实现流域洪水预报记忆单元状态变量知识库建立与预报方案建立。本研究提出的基于知识学习的洪水预报方法的系统理论框架，即“物理机制分析—预报知识描述—预报知识学习—预报知识库建立-预报方案制定与修正”，能够为更有效地进行洪水资源利用提供科学依据并起到重要的指导意义。