Long-term prediction of Poyang Lake water level by combining multi-scale isometric convolution network with quantile regression

The Poyang Lake basin, China Deep learning (DL) models have been widely applied to lake water level (WL) forecasts at short lead times. However, long-term prediction remains a significant challenge. This study introduces a novel multi-scale isometric convolution network (MICN) model to predict long-term lake WL for the first time. Additionally, the quantile regression (QR) method is incorporated into MICN to generate interval forecasts, offering probabilistic ranges of future WL to enhance risk management. The novel MICN model, which comprehensively incorporates both local and global features, successfully achieved robust predictions of Poyang Lake (PYL) WL. In particular, the long-term forecasts (30 days) have been significantly improved, with the Nash-Sutcliffe Efficiency Coefficient (NSE) exceeding 0.98 at each site. Furthermore, the Root Mean Square Error (RMSE) was improved by more than 20 % compared to the next best-performing DL model. Meanwhile, the hybrid QRMICN model demonstrated satisfactory performance in interval prediction for the short-term horizons. The Predicted Interval Coverage Probability (PICP) at four tested stations reached 0.9, while the Predicted Interval Normalized Average Width (PINAW) remained narrow, indicating that the future predicted WL can be well included in the given intervals. This study offers an effective solution for the prediction of PYL WL with highly complex characteristics, which can be also extended to other hydrological variables. • The MICN model was applied to long-term lake WL prediction for the first time. • The MICN model outperforms traditional DL models in lake WL prediction. • The QRMICN model is feasible for short-term lake WL interval prediction.