Evaluating and enhancing the performance of the K-Means clustering algorithm for annual coastal bed evolution applications

The prediction of the coastal bed evolution at an annual scale utilizing process-based models is usually a complex task requiring significant computational resources. To compensate for this, accelerating techniques aiming at reducing the amount of input parameters are often employed. In the framework of this research, a comprehensive evaluation of the capacity of the widely-used K-Means clustering algorithm as a method to obtain representative wave conditions was undertaken. Various enhancements to the algorithm were examined in order to improve model results. The examined tests were implemented in the sandy coastline adjacent to the port of Rethymno, Greece, utilizing an annual dataset of wave characteristics using the model MIKE21 Coupled Model FM. Model performance evaluation was carried out for each test simulation by comparing results to a “brute force” one, containing the bed level changes induced from the annual time series of hourly changing offshore sea state wave characteristics, deeming the results very satisfactory. The best-performing configurations were found to be related to the implementation of a filtering methodology to eliminate low-energy sea states from the dataset. Employment of clustering algorithms utilizing “smart” configurations to improve their performance could become a valuable tool for engineers desiring to obtain an accurate representation of annual bed level evolution, while simultaneously reducing the required computational effort.