Evaluating Hydrological Models in a Regulated River: Implications for Future Water Levels in the Göta River

Sammanfattning

The increased frequency of extreme weather events over the past few years is increasing the risks of natural disasters in landslide-prone regions in Sweden. One often overlooked process in landslide triggering is fluctuations in river stage, which alters the resistance force of the slope. Regulated rivers–a large fraction of the world’s large river systems–do not exhibit typical rainfall-runoff behaviors, as anthropogenic effects complicate those patterns, increasing the difficulties in river stage modeling. However, machine learning, especially deep learning, has become more common in hydrological forecasting, due to its superior pattern-finding abilities, and offers new opportunities for improving predictive modeling of river stage and discharge in regulated rivers. Since machine learning lacks direct physical representation, hybrid modeling, which combines physically based and machine learning approaches, has emerged as an alternative. This study evaluated the performance of three modeling approaches for a regulated river in Sweden: a process-based model (S-HYPE), a deep learning, data-driven approach (LSTM), and a hybrid approach (LSTM as a post-processor to learn the error in the S-HYPE output). Performance of these models for predicting stage and discharge of historical observed data was evaluated, in addition to projecting river stage in a future climate based on two climate models. Since the S-HYPE data was provided as discharge, a streamflow-river stage regression was established to convert discharge to water levels. The results show that modeling river stage in the heavily regulated Göta River is difficult, and further studies with different architectures are needed. However, both the LSTM and Hybrid models outperformed the S-HYPE model, suggesting that machine learning can contribute to improved future predictions. For a future climate, the S-HYPE projects increased extreme water levels in Trollhättan, whilst the LSTM and Hybrid models mainly project an increase in median water levels and suppression of extremes. Additional studies are needed to reduce uncertainties in the impact of climate change on river stage and to clarify their connection to slope stability. Neither of the models managed to capture the high-frequency fluctuations of the hydropower regulations, opening for further studies that could include a top-down approach incorporating operational data if available or by changing the model and trying alternative architectures that have worked well elsewhere.