Optimizing Network Architecture of Artificial Neural Networks (ANNs) in Rainfall-Runoff Modeling
Artificial neural networks (ANNs) are general-purpose techniques that can be used for nonlinear data-driven rainfall-runoff modeling. The fundamental issue to build a worthwhile model by means of ANNs is to recognize their structural features and the diff
- PDF / 342,689 Bytes
- 10 Pages / 439.37 x 666.142 pts Page_size
- 47 Downloads / 210 Views
Abstract Artificial neural networks (ANNs) are general-purpose techniques that can be used for nonlinear data-driven rainfall-runoff modeling. The fundamental issue to build a worthwhile model by means of ANNs is to recognize their structural features and the difficulties related to their construction. Without a doubt, the magnitude and quality of data, the type of noise, and the mathematical properties of the algorithm for estimating the usual large number of parameters are critical for the simplification performances of ANNs. There are many avoiding overfitting techniques in improving the generalization of ANNs. This paper was reported the optimization of the hidden neurons of nonlinear autoregressive with external inputs (NARX) neural network of the tropical river basin. For the Langat River Basin, eight or ten hidden neurons were found to be the most optimal for the ANN model. The next steps should be conducting the optimum numbers of inputs delay to be used for the model since the precise combination of the network architectures features will improve the accuracy level of ANN model.
Keywords Artificial neural networks Hidden neurons Nonlinear autoregressive with external inputs Rainfall-runoff modeling
K. Khalid (&) Faculty of Civil Engineering, Universiti Teknologi MARA Pahang, Jengka, Pahang, Malaysia e-mail: [email protected] M.F. Ali N.F.A. Rahman Faculty of Civil Engineering, Universiti Teknologi MARA, Shah Alam, Selangor, Malaysia M. Husaini Center for Research and Innovation, UniKL Malaysian Spanish Institute, Kulim, Kedah, Malaysia M.R. Mispan Soil & Water Management, MARDI, Serdang, Selangor, Malaysia © Springer Science+Business Media Singapore 2016 M. Yusoff et al. (eds.), InCIEC 2015, DOI 10.1007/978-981-10-0155-0_11
107
108
K. Khalid et al.
1 Introduction Streamflow prediction for a river has been one of the most explored areas in the recent hydrologic study. It is a great concern for the prediction of streamflow with good probability and reliability since the watershed models may show high nonlinearity. The models have been practised may vary from a physical, empirical, and numerical methods and other hybrid black box models for streamflow prediction. The main weaknesses observed in using physical model are the requirement of a more accurate and extensive data set that sometimes is tedious to acquire. The data-driven models may have an advantage within this framework as they require minimum data and may provide satisfactory results. Neural network (NN), genetic algorithm, and fuzzy and hybrid algorithms are some of the methods that have received lots of attention among all modeling techniques during recent decades. In recent years, artificial neural networks (ANNs) have proven to be an efficient alternative to traditional methods for modeling quantitative water resource variables [1, 2]. The ANN concepts and applications in hydrology have been discussed by the ASCE Task Committee on Application of ANN in hydrology, which concludes that ANNs may be perceived as alternative modeli
Data Loading...
