Study On Learning Machine Based Flood Forecast Model

Flood forecasting is an important nonstructural measure of natural disaster prevention and alleviation. The traditional flood forecast methods are relatively complex and hard to be popularized. Since the early nineties last century, machine learning techniques such as artificial neural networks (ANN) have been attempted in flood forecast areas, such as rainfall-runoff modeling and stream flow forecasting, with some valuable experiences achieved. This paper presents several precise, reliable and practical flood forecast models based on some new style learning machines. Their performances were valued in case studies. The components of this paper are listed as follow:(1) The characteristics of Elman recurrent neural networks are introduced. The main course of modeling, including data noise reducing, training controlling and model structure selecting, is proposed. The model's performances, when employing different kinds of input data, have been compared. The performance in case study indicates that flood forecast modeling using Elman ANN is a flexible, extendable and accurately method.(2) The huge size of ANNs is a frustration when they were applied in practice. Principle component analysis (PCA) can be used to reduce the amount of input elements so that the size of ANN will shrink relevantly. Also the training time can be shorten with little lose or even no lose of accuracy.(3) ANNs are often viewed as Black Box models whose parameters don't have any physical meaning. And the structures of ANNs are similar in different hydrologic systems, by this mean, the basic information such as distributing of hydrometric stations can't be utilized. This paper presents a new flood forecast model based on complex ANN, which can make the information of hydrologic systems as guidance when constructing the structure of ANN. The parameters of complex ANN have physical meanings to some extent and the whole model is more reasonable.(4) Support vector machine (SVM) is a novel powerful learning machine, which can solve small-sample learning problem better. The basic ideas of statistical learning theory (SLT) and SVM are introduced, and the characteristics of SVM are illuminated. Meanwhile, the SVM's parameters selection method and the representations of different models are researched.(5) Three types of flood forecast models based on SVM are presented. These models are SVM flood forecast model with changeless training set, dynamic recursion SVM flood forecast model with fixed length training set and dynamic recursion SVM flood forecast model with memory. It is proved to be promising and valuable for these models with experimental results.(6) All these flood forecast models are examined in case studies, and their characteristics are obtained by performances evaluating, which can be considered as guidance in practical use.(7) The concept and principle of pre-evaluation machine (PEM) of flood forecast models are described. PEM models based on resemblance coefficient, ANN and SVM are posed. The feasibility and difficulties of their application are discussed.