Multi-model data fusion for improved streamflow prediction based on hydrological and data-driven models

Sustainable management of water resources relies on accurate river flow prediction. This study explored how multi-model data fusion techniques enhance the reliability of rainfall-runoff modeling by integrating the strengths of process-based and data-driven approaches. Accordingly, we employed different streamflow prediction models, comprising the TUW (Technische Universität Wien) model and the Long-Short-Term Memory (LSTM)-based models, LSTM and Stack-LSTM, to predict streamflow in the Hron River basin in Slovakia during the 2007–2020 time period. An enhanced streamflow prediction system was then developed by merging predictions from multiple models using the Simple Average Method (SAM) and the Bayesian Model Averaging (BMA) approach. The findings revealed that the Stack-LSTM model performed similarly to the LSTM algorithm, and both outperformed the TUW method. Evaluation and analysis showed that the Stack-LSTM model achieved a Nash-Sutcliffe efficiency coefficient (NSE) of 0.98 and a Mean Absolute Percentage Error (MAPE) of 6.96% during the test stage. Furthermore, the comparison of outcomes from the multi-model averaging methods revealed that the BMA approach outperformed the SAM. As a result, the MAPE for the BMA method was reduced by 50.2% compared to the SAM. This research provides a robust tool for streamflow prediction, enhancing decision-making in water resources management.