Advancing water level prediction using clustering-based machine learning techniques in data-scarce regions

Accurate and scalable water level forecasting is essential for effective water resources management, particularly in regions with limited long-term records. We present a clustering-based framework for one- and three-day-ahead water level prediction in the Saemangeum Watershed, South Korea. Twenty-five monitoring stations were grouped into six hydrologically similar clusters using k-means clustering with wavelet-entropy features. Within each cluster, multilayer perceptron (MLP) models were trained using two strategies: (1) training only at the centroid station and (2) training at the station with the longest record in each cluster. The longest-record strategy showed strong agreement with observations, achieving mean Nash–Sutcliffe efficiency and root-mean-square error values of 0.97 and 0.06 for one-day-ahead forecasts, and 0.83 and 0.14 for three-day-ahead forecasts across all stations. By training one MLP per cluster and transferring it to all member stations, the framework reduces computational cost and provides a practical solution for large-scale water level forecasting in data-scarce environments.