Ground subsidence prediction based on a CNN-BiLSTM model integrated with the BKA optimization algorithm

Convolutional neural networks (CNN) and bidirectional long short-term memory (BiLSTM) networks are currently among the most popular deep learning architectures for surface subsidence prediction. However, the selection of hyper parameters for such models remains both time-consuming and complex, and improper choices can significantly degrade model performance. To address this issue, this study integrates the black-winged kite optimization algorithm (BKA) for hyper parameter optimization, constructing a hybrid BKA-CNN-BiLSTM model. Taking Xining city as a case study, experimental analyses were conducted and the results were compared with those of four other models. The findings reveal that the proposed BKA-CNN-BiLSTM model achieves superior performance in both training and prediction. Compared to the BiLSTM model, the coefficient of determination (R²) improvement ranges from 17.43% to 25.77%, and compared to the CNN-BiLSTM model, the improvement ranges from 12.04% to 13.75%. The proposed model also obtains the lowest MAE, MSE, and RMSE among all models. When compared with CNN-BiLSTM models optimized by genetic algorithm (GA) and particle swarm optimization (PSO), the BKA-CNN-BiLSTM model continues to demonstrate higher reliability and prediction accuracy, with R² improvements of 6.20%～17.76% and 1.18%～12.76%, respectively. These results validate the superior performance of the proposed model and offer a novel technical pathway for surface subsidence modeling, while providing practical insights for the application of deep learning in related domains.