What does this research mean for the field?

The proposed AOA-ICEEMDAN-SSA-CNN-BiLSTM ensemble deep learning framework significantly improves runoff forecasting accuracy over benchmark models, and reveals that air temperature is the dominant, seasonally-shifting driver of runoff variability in the Qinghai Lake Basin. Novelty: ClaimNovelty.METHODOLOGICAL. Consensus alignment: ConsensusAlignment.NEUTRAL.

What question did this study set out to answer?

This study aims to improve runoff forecasting accuracy in the Qinghai Lake Basin using a novel ensemble deep learning framework.

May 19, 2026Open Access

An enhanced ensemble deep learning model integrating three strategies for runoff modeling in the Qinghai Lake Basin

Puntos clave

This study aims to improve runoff forecasting accuracy in the Qinghai Lake Basin using a novel ensemble deep learning framework.
Developed AOA-ICEEMDAN-SSA-CNN-BiLSTM framework integrating CNN-BiLSTM, ICEEMDAN decomposition, and dual-optimization.
Utilized historical runoff and meteorological data from 1962–2011 for simulation and forecasting.
Compared performance against benchmark models and analyzed factors influencing runoff variability.
AOA-ICEEMDAN-SSA-CNN-BiLSTM increased R2 by 5.85% compared to BiLSTM.
Reduced RMSE by 16.03%, MAE by 19.05%, and MAPE by 70.89%.
Air temperature identified as the dominant factor affecting runoff variability, with seasonal shifts in its influence.

Resumen

Accurate runoff forecasting is crucial for water resource management in the cold-arid Qinghai Lake Basin (northeastern Tibetan Plateau), yet complex snowmelt dynamics and spatiotemporal heterogeneity limit traditional models. This study develop a novel ensemble deep learning framework, AOA-ICEEMDAN-SSA-CNN-BiLSTM, aiming to enhance simulation stability and precision for hydrological processes during 1962–2011 in Qinghai Lake Basin. The framework couples three strategies: (i) A CNN-BiLSTM hybrid architecture for concurrent spatial feature extraction and temporal sequence learning, (ii) ICEEMDAN (Improved Complete Ensemble Empirical Mode Decomposition with Adaptive Noise) decomposition to resolve non-stationary runoff signals into trainable components, and (iii) A dual-optimization framework utilizing the Arithmetic Optimization Algorithm (AOA) to tune ICEEMDAN parameters and the Sparrow Search Algorithm (SSA) for CNN-BiLSTM hyperparameter selection. Based on historical runoff and meteorological variables, the proposed framework demonstrates superior simulation and forecasting performance compared with benchmark models. Specifically, compared with the BiLSTM model, the proposed framework increases the coefficient of determination (R2) by 5.85%, while reducing the root mean square error (RMSE), mean absolute error (MAE), and mean absolute percentage error (MAPE) by 16.03%, 19.05%, and 70.89%, respectively. Furthermore, explainable learning analysis by SHapley Additive exPlanations (SHAP) reveals that air temperature is the dominant factor influencing runoff variability, followed by precipitation and evaporation, and its contribution shifts seasonally from negative during the frozen period to positive during the snowmelt season. These findings demonstrate the effectiveness of the proposed framework and provide valuable insights for hydrological risk assessment and basin-scale water-resources management.

Me gusta

Guardar

Ver artículo completo