Investigation of the performance of neural gas networks in hydrological clustering

The design of many infrastructures and construction projects requires the extensive studies in the area's geographical conditions and climatic characteristics. The effectiveness of this research itself depends on the information and data required. In many cases, the project area is in a situation where no climatic information such as rainfall is available. Hence, regional frequency analysis has been received much attention. In this way, having specific conditions and mechanisms, the available information in other sites can be expanded and transferred to the other areas. In this research, clustering is one of the most effective steps that divide the existing  stations into the hydrologically  homogeneous areas. Therefore, in this study, in addition to the common methods in clustering, two new models, neural gas network and growing neural gas network, were used to determine the homogeneous regions in Khuzestan province, Iran. One of the unique features of these algorithms is learning the topology or shape of the distributions that governs the data space. Using the variables of longitude, latitude, altitude, mean annual rainfall, and maximum 24-hour rainfall of the station, the design area was divided into two hydrologically homogeneous areas, and the clustering process was performed. The results show that the neural gas networks has a high efficiency and accuracy in view of clustering. The Mean Percentage Difference and Coefficient of Variation of Root Mean Square Error in neural gas were estimated to be 15.56 and 24.39 percent, respectively, which showed a considerable advantages over the conventional methods.