深度学习在水文预测领域的应用进展与思考
2021-08-10陈爱青
陈爱青
摘 要:合理的水文预测模型是水文水资源决策管理的基础,如何充分挖掘既有水文数据中的信息成为当前水文预测领域的一大挑战,深度学习方法的快速发展为水文预测提供了新的思路。针对国内外近期提出的水文预测深度学习模型进行归纳总结,从数据来源、方法模型、验证讨论等方面探讨了深度学习在水文预测领域的应用进展,并进一步给出了数据获取、模型迁移、实时预警等方面的思考。结果表明:充足的水文数据是精准预测的前提,合理的模型构建策略是考虑各种不确定因素的关键手段,模型的适用性和实时预警是未来进一步研究的方向。
关键词:水文预测 深度学习 大数据 长短时记忆神经网络 卷积神经网络
中图分类号:P338 文献标识码:A 文章编号:1674-098X(2021)04(c)-0252-05
Application Progress and Thinking of Deep Learning in the Field of Hydrological Prediction
CHEN Aiqing
(Shanghai Branch of Changjiang Survey Planning Design Research Co., Ltd., Shanghai, 200439 China)
Abstract: A reasonable hydrological prediction model is the foundation of hydrological and water resources decision-making and management. How to fully mining the information in existing hydrological data has become a major challenge in the current hydrological prediction field. The rapid development of deep learning methods provides new ideas for hydrological prediction. In view of the recent domestic and foreign hydrological prediction deep learning models, this paper discusses the application progress of deep learning in the field of hydrological prediction from the aspects of data sources, proposed models, verification and discussion, and further provides some suggestions on data acquisition, model migration, and real-time early warning. The results show that sufficient hydrological data is a prerequisite for accurate prediction, a reasonable model building strategy is a key means to consider various uncertain factors, and the applicability of the model and real-time early warning are the directions for further research in the future.
Key Words: Hydrological prediction model; Deep learning; Big data; LSTM; CNN
水文模型是水文水資源管理、防汛减灾的基础手段,传统的概念性模型、物理模型和统计学模型已经在径流、降水量预测等领域取得了长足进展,但随着传感手段的不断发展,海量水文数据的积累导致传统方法难以充分挖掘现有水文数据中的有效信息,近年来深度学习方法以其处理高维度、多特征海量数据方面取得了显著成果,已经被广泛应用于水文预测 领域。
作为机器学习的一种,深度学习概念是Hinton 于2006年在《Science》发表的“深度信念网络(Deep belief network,DBN)”,他将多层感知机(多层神经网络)的学习方法称为“深度学习”。简单来说,深度神经网络可以理解为具有2个或2个以上隐藏层的多层人工神经网络,深度学习是相对于浅层学习而言的,深度学习通过更深的网络结构来实现更好的特征表示。……
