Quantitative Analysis of Input Schemes and Key Variable Contributions in River Runoff Forecasting Models

In Long Short-Term Memory (LSTM)-based runoff forecasting models, the selection of input schemes is critically important. This study, using daily rainfall and runoff data from the Jingle Basin (2006–2014), investigated three input schemes to evaluate their forecasting efficacy and employed the Shapley Additive Explanation (SHAP) method to quantitatively analyze variable contributions. The results demonstrate that LSTM model performance deteriorates with increasing lead time, achieving optimal accuracy at a 1-day lead (MAE: 0.90 m 3 /s, RMSE: 3.09 m 3 /s, NSE: 0.84). The results, validated by significance testing, are reasonable; incorporating precipitation characteristics significantly enhances model performance compared to baseline schemes, reducing RMSE by 6–34% and improving NSE by 9–14%. SHAP analysis reveals antecedent runoff as the dominant influencing factor, accounting for 65.9–84.7% of total importance. Furthermore, the contributions of trend, seasonal, and residual components progressively increase with extended lead times, demonstrating non-negligible roles in forecast outcomes. These findings, confirmed by significance testing, provide quantitative insights into input variable contributions to target uncertainty and enhance the mechanistic understanding of precipitation-runoff relationships, offering valuable references for optimizing hydrological forecasting systems.