Abstract: Extreme meteorological events have been occurring frequently worldwide, posing dual challenges to renewable energy integration and secure power supply for the grid systems with a high hydropower proportion. To enhance the systems’ dynamic responding capability to such extreme events, this study describes a new methodology for generating and selecting the extreme streamflow scenarios of a river basin by coupling its key meteorological and hydrological elements. First, we use sensitivity analysis based on the Shapley additive explanations (SHAP) theory to reveal the critical influence of precipitation, soil moisture content, and air temperature over the basin on its streamflow. Then, an adaptive machine learning framework by coupling meteorological-hydrological elements with the streamflow, is constructed. It uses a Markov Chain Monte Carlo (MCMC) approach to simulate the extreme meteorological-hydrological events, and integrates the historical observational data as inputs to generate an ensemble of the streamflow scenarios. Finally, the scenarios are integrated using an enhanced K-means clustering algorithm, and the dissimilarity between individual scenarios within each cluster to the cluster centroid is calculated by combining with a modified Dynamic Time Warping (DTW) algorithm to select the optimized extreme streamflow scenarios based on the principle of maximum dissimilarity. Our method proves effective and applicable through validation using the streamflow data (1952-2006) from the Wujiang River basin in Southwest China and the corresponding meteorological-hydrological records.

Key words: extreme runoff scenarios, machine learning algorithm, Markov chain, Shapley additive explanations theory