Classification of reservoir flood processes based on self-organizing map neural networks

In this study, self-organizing map (SOM) neural networks were constructed based on entire flood processes for reducing information loss and subjectivity compared to traditional flood classification methods. Considering evaluation indicators of representativeness and topology, we identified the optimal dimensionality of the SOM for characterizing the topological logic relationship of flood processes and realizing a new flood classification. The inflow flood process of the Three Gorges Reservoir (TGR), located along the main stream of the Yangtze River, served as a study case. Our results showed the following. ① The coverage ratio indicator and the gap of a 2×3-dimensional SOM were 56.7% and 2%, respectively, of those of a 3×3-dimensional SOM; besides, the former SOM had only one flip and its topology was better than that of the latter. Due to the good performance of the 2×3-dimensional SOM, this was identified as the most suitable for the classification of TGR flood processes. ② Based on the 2×3-dimensional SOM, the flood processes registered at the Yichang station were divided into 6 classes. The differences between classes were reflected by neuron topology, indicating that this SOM can provide reliable flood classification results from high-dimensional flood data by visualizing the neuron topology. ③ Compared with traditional classification methods, which have been applied to investigate historical typical floods, the proposed 2×3-dimensional SOM was found to be more reliable and to provide more useful information.