Elsevier

Journal of Hydrology

Volume 375, Issues 3–4, 15 September 2009, Pages 450-458
Journal of Hydrology

A hybrid neural network model for typhoon-rainfall forecasting

https://doi.org/10.1016/j.jhydrol.2009.06.047Get rights and content

Summary

A hybrid neural network model is proposed in this paper to forecast the typhoon rainfall. Two different types of artificial neural networks, the self-organizing map (SOM) and the multilayer perceptron network (MLPN), are combined to develop the proposed model. In the proposed model, a data analysis technique is developed based on the SOM, which can perform cluster analysis and discrimination analysis in one step. The MLPN is used as the nonlinear regression technique to construct the relationship between the input and output data. First, the input data are analyzed using a SOM-based data analysis technique. Through the SOM-based data analysis technique, input data with different properties are first divided into distinct clusters, which can help the multivariate nonlinear regression of each cluster. Additionally, the topological relationships among data are discovered from which more insight into the typhoon-rainfall process can be revealed. Then, for each cluster, the individual relationship between the input and output data is constructed by a specific MLPN. For evaluating the forecasting performance of the proposed model, an application is conducted. The proposed model is applied to the Tanshui River Basin to forecast the typhoon rainfall. The results show that the proposed model can forecast more precisely than the model developed by the conventional neural network approach.

Introduction

Taiwan is situated in one of the main paths of northwestern Pacific typhoons. There are about four typhoons invading Taiwan each year. During the typhoon season (generally from July to October), heavy rainfall often causes floods and results in loss of life and property damage. Precise forecasts of typhoon rainfall are needed for issuing flood warnings. Therefore, to improve the accuracy of typhoon-rainfall forecasts is always an important task of flood management in Taiwan. However, typhoon rainfall is a complex component of the hydrologic cycle. It is not easy to develop a physically based model for typhoon-rainfall forecasting because of the lack of knowledge about its physical process. It is also difficult to construct a statistically based model using traditional regression techniques owing to various factors involved and their great variability in space and time.

Artificial neural networks (ANNs) have been proven to be very successful in dealing with highly-complicated problems. Due to their powerful capability to model nonlinear systems without the need to make any assumptions, ANNs have found increasing applications for modeling hydrological processes. A review of ANNs in hydrology was presented by the ASCE Task Committee on Application of Artificial Neural Networks in Hydrology, 2000a, ASCE Task Committee on Application of Artificial Neural Networks in Hydrology, 2000b. Additionally, some studies were performed to examine the potential of combining ANNs with other computing algorithms (Lin and Chen, 2005b, Wu and Chau, 2006, Muttil and Chau, 2006, Xie et al., 2006). Various applications of ANNs for hydrologic forecasting (Lin and Chen, 2004a, Ramirez et al., 2005, Pulido-Calvo and Portela, 2007, Partal and Cigizoglu, 2008) were studied in recent years. As regards the rainfall forecasting by ANNs, the multilayer perceptron network (MLPN) is the most commonly used (French et al., 1992, Luk et al., 2000, Luk et al., 2001, Lin and Chen, 2005a). By using the back-propagation algorithm to network parameter estimation, the MLPN can approximate complicated nonlinear functions to arbitrary accuracy (Govindaraju and Rao, 2000). Hence, the MLPN is used herein to construct the nonlinear relationship between the influential factors and the typhoon-rainfall depth.

However, to determine the appropriate model inputs is an important task when the MLPN is used for forecasting. If the appropriate inputs are used, the performance of ANN model can be improved (Bowden et al., 2005, Lin and Chen, 2008). Thus, the determination of the appropriate model inputs is also considered in this paper. For rainfall forecasting, Luk et al. (2000) used the spatial rainfall information from nearby rain gauges as input data and indicated that the inclusion of spatial information improves the accuracy of forecasts. But it is not easy to determine the optimal number of nearby rain gauges, especially in the situation that there are many rain gauges in the study area. A geostatistical analysis method, the semivariogram, is applied to help the determination of the optimal number of nearby rain gauges (Lin and Chen, 2005a). In addition to the spatial rainfall information, the typhoon characteristics (namely, the latitude and longitude of the typhoon center, the maximum wind speed near the center, the atmospheric pressure of the center, the radius of winds over 15 m/s, the typhoon moving speed, the typhoon moving direction, and the distance of the typhoon center from a rain gauge) also affect the rainfall during a typhoon event. Although these characteristics are too complex to model directly, they are still the helpful information to typhoon-rainfall forecasting. Hence, both the typhoon characteristics and the spatial rainfall information from nearby rain gauges are used as input to the MLPN. However, the input of MLPN becomes high-dimensional and complex. If used carelessly, the MLPN easily learns irrelevant information, which will reduce the performance of forecasting. In order to improve the typhoon-rainfall forecasting, one needs a data analysis method to analyze the input data before the development of the MLPN.

Self organizing map (SOM) proposed by Kohonen is a special class of ANN. It can project high-dimensional input space on a low-dimensional topology to demonstrate the topological relationships among input patterns, which helps users to discover the information. Hence, SOM can provide features that facilitate insight into the hydrological processes (Hsu et al., 2002). The property has rendered the SOM an attractive alternative for clustering massive and complex data which traditionally have been the domain of conventional clustering methods. The superior performance of SOM in clustering has been discussed in the literature (Mangiameli et al., 1996, Michaelides et al., 2001, Lin and Chen, 2006, Lin and Wu, 2007). In addition to cluster analysis, discrimination analysis can also be performed by the identical SOM (Lin and Wang, 2006). Lin and Wang (2006) indicate that SOM has great potential for developing a data analysis technique. Hence, in this paper, the SOM is adopted to develop a data analysis method which in turn helps the development of MLPN. There are two advantages of using SOM to develop the data analysis technique. First, the SOM-based data analysis technique can cluster data without specifying the number of clusters in advance. This property helps users determine the number of clusters more objectively. Second, the SOM-based data analysis technique can integrate cluster analysis with discrimination analysis together, which allows users to obtain the information more directly.

The objective of this paper is to propose a hybrid neural network model which combines the SOM with the MLPNs to improve the accuracy of typhoon-rainfall forecasts. With the advantages of the data analysis technique developed by SOM and the capability of nonlinear function approximation of MLPNs, the proposed model is expected to be useful for typhoon-rainfall forecasting. The performance of the proposed model is examined by actual applications to the Tanshui River Basin in northern Taiwan. In addition, the proposed model is compared with a forecasting model developed by the conventional neural network approach (hereafter referred to as the conventional model). In contrast to the proposed model, the conventional model is constructed by only one MLPN which uses non-grouped data as input.

Section snippets

Methodology

The proposed model is illustrated in Fig. 1. The basic concept of the proposed model is that data with significant different properties should be separated into different clusters before the nonlinear regression is developed by MLPN. In other words, the proposed model can be regarded as a piecewise nonlinear multivariable regression model. Hence, a data analysis technique is developed herein to perform cluster analysis as well as discrimination analysis of the input data. For each cluster, the

The study area and data

The study area of this paper is the Tanshui River Basin located in northern Taiwan. The basin has an area of 2726 km2 and an average river slope of 1/45. The average annual rainfall of the basin is 2966 mm, 63.5% of which occurs during the wet season (from May to October). The study area and the locations of rain gauges are shown in Fig. 3. In this basin, heavy typhoon rainfall often causes serious natural hazards such as floods and debris flow. Typhoon-rainfall forecasting can extend the time to

Results and discussions

The comparison test on the forecast accuracy is made among three MLPNs which use different time lags of typhoon characteristics as input. The three MLPNs are formulated asr(t+1)=f1[Ty(t)]forMLPN-1r(t+1)=f2[Ty(t),Ty(t-1)]forMLPN-2r(t+1)=f3[Ty(t),Ty(t-1),Ty(t-2)]forMLPN-3where the fi [ ] represents the ith MLPN, r(t + 1) is the rainfall depth at time t + 1, and Ty(t), Ty(t  1) and Ty(t  2) are the typhoon characteristics at time t, t  1 and t  2, respectively. Table 1 summarizes the results of the

Summary and conclusions

A hybrid artificial neural network model which combines the SOM with the MLPNs is proposed in this paper to forecast the typhoon rainfall. Firstly, the input determination is performed. By using a comparison test, the present and the past one-hour typhoon characteristics are decided as the optimal input factors. In addition, the semivariogram method is used to help determine the required number of nearby rain gauges whose rainfall records are used as the spatial information and involved in

References (28)

  • J.E. Nash et al.

    River flow forecasting through conceptual models. Part 1: a discussion of principles

    Journal of Hydrology

    (1970)
  • T. Partal et al.

    Estimation and forecasting of daily suspended sediment data using wavelet-neural networks

    Journal of Hydrology

    (2008)
  • I. Pulido-Calvo et al.

    Application of neural approaches to one-step daily flow forecasting in Portuguese watersheds

    Journal of Hydrology

    (2007)
  • ASCE Task Committee on Application of Artificial Neural Networks in Hydrology

    Artificial neural networks in hydrology I: preliminary concepts

    Journal of Hydrologic Engineering

    (2000)
  • Cited by (92)

    View all citing articles on Scopus
    View full text