Abstract: Accurate river inflow forecasting at cross-sections is crucial for water resource management, flood warning and disaster mitigation, ecological protection, and hydropower project planning.To improve the prediction accuracy of a single river inflow forecasting model, a forecasting method(LASSO-QRLSTM) is proposed by combining LASSO regression, Quantile regression long short-term memory(QRLSTM) networks, and kernel density estimation(KDE).First, LASSO regression is applied to extract key explanatory variables from high-dimensional river inflow feature vectors, reducing the complexity of the nonlinear relationship between explanatory and response variables.Next, a QRLSTM model is established for river inflow forecasting to obtain quantile predictions at various points.KDE is then used to fit the probability density function, providing possible future river inflow values and their associated probabilities, leading to the final forecast.The proposed model is applied to river inflow forecasting at key sections of the Xijiang River and Gaoyao Hydrological Station in Guangdong Province, and compared with LASSO-QRNN,LASSO-GBDT,QRLSTM,QRNN,and GBDT models.The results show that:(1) The hybrid model combining LASSO regression outperforms the single QRLSTM,QRNN,and GBDT models.(2) In particular, the LASSO-QRLSTM model achieves anRMSEof 1 804.270 m3/s and anNSEof 0.973 for the river inflow forecasting at the Sixianjiao Section.In terms of probabilistic indicators, the LASSO-QRLSTM model yields the best results, with a Continuous Ranked Probability Score(CRPS) of 842.618 and a Pinball Loss(PL) value of 465.964.The model exhibits superior prediction performance, especially in extreme values, with better fitting and narrower prediction intervals.(3) Further validation on Gaoyao Hydrological Station confirms its adaptability and stability.The research findings can provide valuable insights for accurate hydrological forecasting and optimal allocation of water resources.