Streamflow prediction using LASSO-FCM-DBN approach based on hydro-meteorological condition classification

•The LASSO-FCM-DBN approach can improve the performance of streamflow prediction.•The performance of the FCM classification is more stable than the threshold method.•DBNs performed better than traditional ANNs in all three statistical measures considered. Streamflow prediction is a challenging task due to the different processes involved in streamflow generation. These different processes have different characteristics of the relationships between hydro-meteorological variables and streamflow, which make it a challenging task to develop single data-driven stream flow prediction models that can map the input-output relationships for all different streamflow regimes. To improve the performance of streamflow prediction, we proposed a flow-regime-dependent approach to map the relationships between hydro-meteorological variables and streamflow based on hydro-meteorological condition classification. This approach integrates the least absolute shrinkage and selection operator (LASSO), Fuzzy C-means (FCM) and Deep Belief Networks (DBN) and therefore referred to as the LASSO-FCM-DBN approach. This approach employs LASSO to select the hydro-meteorological variables which have a significant impact on streamflow, FCM to identify different streamflow regimes, and DBN as a data-driven model to map the nonlinear and complex relationships between the selected hydro-meteorological variables and streamflow within different flow regimes. To assess the performance of the proposed approach, two comparative studies were carried out – 1) the multi-variable FCM was compared to the traditional single-variable threshold-based method; and 2) the performance of the DBN was compared to a traditional Artificial Neural Networks (ANNs) model. Two stations in the Tennessee River, USA were used as the case study. The results demonstrate that the performance of the multi-variable-based FCM classification method is better and more stable than the traditional threshold-based single-variable method, due to the sensitivity of the single-variable method to different threshold values. In addition, DBNs performed better than traditional ANNs in all three statistical measures considered. Overall, the LASSO-FCM-DBN multi-model system significantly improved the performance of streamflow prediction and is therefore a valuable tool for water resources management and planning.