Identification of support vector machines for runoff modelling

Abstract

This paper describes an exploration in using SVM (Support Vector Machine) models, which were initially developed in the Machine Learning community, in flood forecasting, with the focus on the identification of a suitable model structure and its relevant parameters for rainfall runoff modelling. SVM has been applied in many fields and has a high success rate in classification tasks such as pattern recognition, OCR, etc. The applications of SVM in regression of time series are relatively new and they are more problematic in comparison with classifications. This study found that exhaustive search of an optimum model structure and its parameter space is prohibitive due to their sheer size and unknown characteristics. Some parameters are very sensitive and can increase the CPU load tremendously (and hence result in very long computation times). All these make it very difficult to efficiently identify SVM models, which has been carried out by manual operations in all study cases so far. The paper further explored the relationships among various model structures (ξ-SV or ν-SV regression), kernel functions (linear, polynomial, radial basis and sigmoid), scaling factor, model parameters (cost C, epsilon) and composition of input vectors. These relationships should be able to provide useful information for more effective model identification in the future. The unit response curve from SVM was compared with a transfer function model and it is found that a TF model outperforms SVM in short-range predictions. It is still unclear how the unit response curve could be utilised for model identification processes and future exploration in this area is needed.