基于数据驱动的CNN洪水演进预测方法

洪水是对人类生命财产造成威胁的重大自然灾害之一，使用洪水演进计算结果来提高抗洪救灾应急的效果是一种行之有效的重要方法。现有的洪水演进计算主要基于二维水力物理模型，其模型建立和计算时间都比较长，不能满足快速应急响应的时间要求。本文使用水力模型数值模拟结果作为数据驱动，构建了卷积神经网络（Convolutional Neural Networks，CNN）模型，将训练的CNN模型应用于洪水演进的预测计算。结果表明：CNN预测模型计算时间大幅缩减，大大提高了洪水演进的计算效率；模型可计算1 ~ 6 h的洪水演进，1 ~ 2 h预测结果与水力模型数值模拟相近，总体满足防灾减灾需要；基于数据驱动的CNN预测方法为洪水演进计算问题提供了深度学习的新途径和方法，也可用于任何有输入和输出判断值和目标值和的科学与工程问题。

CNN flood routing method based on data-driven training

Flooding is one of the severest natural disasters to human lives and properties; to reduce flood emergency response time, an efficient vital method is to use flood routing results. Most of the previous flood routing calculations are based on the traditional 2D-hydro-physical model that requires massive computer capability and high CPU cost, failing to meet the demand by quick emergency response. This paper uses the results of a 2D-hydro-physical model as the driving data to train a CNN model, and then this CNN model is applied to flood routing. Results show that the new model brings about a huge reduction in CPU cost and promotes dramatically the efficiency of flood emergency response in reality. It can calculate flood routings of 1-6 hours long, a duration much longer than the traditional one of 1-2 hours, and its overall results meet engineering demands. Thus, the data-driven CNN method is a new approach and methodology for flood routing and useful for other science and engineering problems with specific inputs and outputs.