Hybrid STL-SARIMA forecasting of reservoir inflows in climate-vulnerable basins: a case study in the Yalong River

Climate-induced hydrological non-stationarity (e.g., intensified drought-flood transitions) challenges inflow forecasting in climate-vulnerable basins like the Yalong River, thereby constraining efficient water resources management. Given the non-stationary and periodic characteristics of the runoff series, this study proposes a novel hybrid forecasting model, named STL-SARIMA, which hybridizes Seasonal-Trend decomposition using Loess (STL) with the Seasonal Autoregressive Integrated Moving Average (SARIMA) model, observed runoff data from the Ertan Hydropower Station for the period 2008–2013 were collected. Based on the Seasonal-Trend decomposition procedure using Loess (STL) method, the original data were decomposed into trend, seasonal, and residual components. Combined forecast values for future runoff were then obtained by integrating the features of these sub-series. Finally, the prediction results were compared with those from single models, namely the Autoregressive Integrated Moving Average (ARIMA) and Seasonal Autoregressive Integrated Moving Average (SARIMA). The results show: The hybrid model integrating time series decomposition and SARIMA achieved a Root Mean Square Error (RMSE) of 1,374.07, demonstrating a 6.06% reduction in error compared to the standalone SARIMA model and a 17.45% reduction relative to the standalone ARIMA model. During the prediction process, an exhaustive search optimization method is employed to determine model parameters (2,160 combinations), while the enhancement effects of seasonal periodic components in the data and normalization of raw input data on prediction accuracy were investigated. This study establishes scientific support for predicting runoff in hydrologically abundant yet climatically vulnerable basins.