Enhancing Runoff Forecast Performance in the Yellow River Source Region: Roles of Meteorological Factors, Future Information, and Spatial Differentiation

The Source Area of the Yellow River (YRSA) serves as a critical water conservation region for the Yellow River Basin, and its hydrological processes exert profound impacts on basin-wide water resources management. However, the runoff yielding mechanisms in this region remain insufficiently understood, the underlying surface conditions are highly complex, and the monitoring infrastructure is relatively sparse, together posing substantial challenges to accurate runoff forecasting. Focusing on the Jungong (JG) hydrological station in the YRSA, this study develops a runoff forecasting model based on a Back-Propagation Neural Network (BPNN). Experiments are conducted at both daily and ten-day timescales during the flood and non-flood seasons to systematically investigate the impacts of meteorological factors, future meteorological information, and spatial heterogeneity on forecasting accuracy. The results are compared against the practical forecasting model, namely Multiple Linear Regression (MLR), currently used by the Hydrology Bureau of Yellow River Conservancy Commission (YRCC). The findings indicate that: (1) Incorporating meteorological factors significantly improves forecast accuracy. Using a three-factor combination of precipitation, air temperature, and runoff as inputs increased the Deterministic Coefficient (DC) by 14.3%, 109.8%, and 35.2% at the daily timescale during flood season, ten-day timescale during flood season, and ten-day timescale during non-flood season, respectively, compared with using runoff alone. (2) Considering future meteorological information further enhances runoff forecasting performance, with DC values at the forecast end of the three-factor combination model increasing by 5.5%, 149.3%, and 23.8%, respectively, compared to without consideration. (3) After subdividing the entire study area according to natural geographical characteristics and considering the spatial heterogeneity of meteorological data, the DC values at the forecast end can be further improved to 0.85, 0.70, and 0.83, respectively. (4) Compared with the MLR model used by the Hydrology Bureau of YRCC, the BPNN model developed in this study demonstrates superior performance in most scenarios and provides methodological references and technical support for practical runoff forecasting.