Mixture of experts leveraging Informer and LSTM variants for enhanced daily streamflow forecasting

Abstract

Streamflow forecasting is of paramount importance for water resources management and flood prevention. Machine learning, particularly deep learning, has had significant success in hydrological forecasting. However, there is still a desire for newer single-type and integrated architectures to further enhance the accuracy and reliability of forecasts. Recently, Transformer-based models have emerged as promising tools, and their effectiveness in streamflow modeling tasks warrants further investigation. The Mixture of Experts (MoE) model has also demonstrated potential in other fields, but its application in the hydrological domain remains relatively limited. This study presents an innovative streamflow forecasting model for the Quinebaug River Basin in Connecticut, USA, based on the MoE framework. Firstly, the hyperparameters of expert models, including LSTM, GRU, LSTM-Sequence to Sequence-Attention, and Informer, with lead times ranging from 1 to 8 days, are optimized using the grid search method. Subsequently, Random Forest, LSTM, and Transformer are used as routers to construct 4-class and 2-class MoE frameworks. Finally, the classified outputs are integrated to synthesize the streamflow forecasting results. The results indicate that the Informer model outperforms other benchmark models in all forecast periods, especially in shorter ones. Both 4-class and 2-class MoE can improve the streamflow forecasting results of the optimal sub-model to some extent: when the lead time reaches 5 days or more, the NSE increases by nearly or more than 3 %. This study highlights that the MoE framework can improve daily streamflow forecasting accuracy by integrating the strengths of different experts, although the router tends to prioritize models with superior performance during the classification process.