Groundwater level prediction based on SSA-optimized self-attention mechanism and BiLSTM hybrid model

Xiaolangdi North Bank Irrigation District, Jiyuan City, Henan Province, China. This work focuses on groundwater level prediction in Xiaolangdi North Bank Irrigation District, Henan Province, using the Sparrow Search Algorithm-optimized CNN-BiLSTM-Attention (SSA-CBLA) model. It pursues two primary objectives: (1) to develop a novel integrated forecasting framework that synergizes data augmentation, convolutional and recurrent networks, and an attention mechanism, specifically designed to address the challenges of data scarcity and complex hydrological dynamics; and (2) to implement a systematic hyperparameter optimization strategy using the Sparrow Search Algorithm (SSA) to enhance model efficacy and robustness. The methodology includes circular translation, Gaussian noise injection, spatial-temporal feature modeling and global optimization, validated through model comparisons and ablation experiment. The SSA-CBLA model excels in groundwater level prediction here, with RMSE= 0.0958 and R² = 0.9642. Convolutional Neural Network (CNN) and Bidirectional Long Short-Term Memory (BiLSTM) complementarily boost feature expression. Self-attention cuts redundant info, and data augmentation eases small-sample overfitting. The model is robust to noise (3 %) and window size 12. Errors accumulate with prediction span, and input features are limited. Integrating more geospatial, meteorological and human activity data is needed. • Develop a hybrid model to predict groundwater levels in research area. • Use data augmentation to mitigate overfitting in small-sample hydrological scenarios. • The model has low prediction error and outperforms 18 comparison models. • CNN, BiLSTM and self-attention collaborate to enhance feature extraction. • SSA optimizes model parameters to support regional water resource management.