Cluster Based Pattern Recognition and Analysis of Streamflows

Traditional methods of streamflow analysis and synthesis are based upon information contained in individual data. These methods ignore information contained in and among groups of data. Recently, the concept of extracting information from data groupings through pattern recognition techniques has been found useful in hydrology. For streamflow analysis, this paper proposes several objective functions to minimize the classification error encountered in currently used techniques employing minimum Euclidean distance. The relevance of these functions has been tested on the streamflow data at the Thames river at Thamesville. Specifically, three objective functions considering the properties of shape, peak, and gradient of streamflow pattern vectors are suggested. Similar objective functions can be formulated to consider other specific properties of streamflow patterns. AIC, infra and inter distance criteria are reasonable to arrive at an optimal number of clusters for a set of streamflow patterns. The random initialization technique for the K-mean algorithm appears superior, especially when one is able to reduce initialization runs by 20 times to arrive at optimal cluster structure. The streamflow synthesis model is adequate in preserving the essential properties of historical streamflows. However, additional experiments are needed to further examine the utility of the proposed synthesis model.