Intelligent Prediction of Autumn Flood Season Runoff in Danjiangkou Reservoir Based on TabPFN Model

To improve the runoff prediction accuracy of Danjiangkou Reservoir during the autumn flood season and support the water resource regulation and flood control safety of the water source area of the Middle Route of the South-to-North Water Diversion Project, a two-stage intelligent prediction technical framework of "meteorological factor bias correction - deep learning model construction" was proposed. Firstly, aiming at the systematic bias of ECMWF meteorological forecast data, the XGBoost-SHAP model was used to screen 60 key teleconnection factors (atmospheric circulation factors accounting for 57%). Combined with the Bayesian-optimized XGBoost, LightGBM and Random Forest ensemble tree models, bias correction was carried out. The XGBoost model showed the optimal correction effect (NSE = 0.94 in the training period and NSE = 0.79 in the testing period). Secondly, a Bayesian-optimized bidirectional long short-term memory neural network (BO-BiLSTM) model was constructed for runoff prediction, and a small-sample deep learning model TabPFN (pre-trained based on Transformer architecture) was mainly introduced to solve the problem of insufficient generalization ability caused by the short data sequence in the upper reaches of the Hanjiang River. The results show that the TabPFN model achieves an NSE of 0.74 and a qualification rate of 67.96% in the testing period with a 1-month lead time, which is 16.02% higher than that of the BO-BiLSTM model at the same lead time. The qualification rate remains between 56.20% and 63.80% for the long lead time of 2-6 months, providing key technical reference for autumn flood season runoff prediction in small-sample watersheds.