Exploring Extreme Flow Events and Associated Patterns in Switzerland: a Dense Feed-Forward Neural Network Approach

Switzerland relies significantly on sustainable water management to meet its diverse socio-economic and environmental needs. As climate change introduces heightened uncertainty in weather patterns, accurate forecasting of extreme flows from climatic data has become essential for efficient water resource management in the country. Furthermore, these events are likely shaped by nonlinear hydroclimatic and compound conditions distinct from typical average cases. A thorough understanding of these phenomena is therefore crucial for effective adaptation to changing climatic conditions. In this regard, data-driven techniques, such as Machine Learning algorithms, have proven capable of extracting knowledge from vast amounts of data, providing valuable insights into the underlying climate and societal dynamics driving extreme flow events. The aim of the present study is therefore twofold. First, we evaluate the ability of a Dense feed-forward Neural Network (DNN) model to predict drought and peak flow events in Switzerland based on anthropogenic, environmental and climatic data. On the other side, we investigate the role of each driver in the prediction and we study the temporal trends of the target and the features. The analysis was conducted on a large dataset consisting of daily discharge data from more than 400 sites across the country, from 1999 to 2019. First, we evaluated the flow distribution at each individual site, considering only the extreme events and developing two distinct DNN models for droughts and for peaks. The DNN performed better in modeling droughts, achieving in the test set a mean Nash-Sutcliffe efficiency coefficient of 0.6 and a mean Kling-Gupta efficiency coefficient of 0.8, compared to 0.1 and 0.38, respectively, for the peaks. A sensitivity analysis of the features, such as the cumulative precipitation and mean air temperature in the preceding weeks of the event, was performed. In addition, we delved into a detailed examination of the temporal trends of the climatic drivers and the extreme flow rates over the 20 years of the study. In the subsequent phase of the project, we explored a multi-site modeling approach to address the issue of the DNN model's poor performance in predicting peak flows. We introduced geographic, land use and other anthropogenic factors specific to each watershed. By revealing the predictive potential of data-driven models, this study serves as a valuable foundation and resource for addressing extreme flow events and the hydroclimatic and anthropogenic patterns behind them.