A combined prediction model for middle and long-term water and sediment forecasting based on causal analysis feature selection

Study region: Upper reaches of the Weigan River Basin, southern foothills of the Tianshan Mountains, Xinjiang, China Study focus: The stochastic characteristics of water-sediment fluxes in the basin pose significant challenges for accurate mid-to-long term predictions. In complex hydrological systems, collected multivariate time series contain critical system features, yet traditional feature selection methods fail to consider inter-variable causal relationships. This study proposes an ensemble prediction model incorporating causal feature selection, developing 24 individual prediction models for water-sediment processes and selecting optimal combinations to enhance prediction accuracy for reservoir inflow water and sediment. New hydrological insights for the region: By employing causal analysis instead of conventional feature selection and applying SHAP for LSTM gradient interpretation, we identify key causal variables: historical runoff, temperature, and evapotranspiration primarily govern reservoir inflow (with snowmelt being the dominant water source), while runoff, historical sediment load, and Nino1+2 most influence sediment inflow. For higher-quality runoff series, GRMIG-MCI-based models achieve KGE>0.85, whereas PCMCI proves more suitable for noisy sediment series (KGE≈0.4). The monthly-weighted ensemble approach improves NSE by 0.35%–3.65% for runoff and over 20% for sediment compared to individual optimal models.