Optimization of Hidden Size Parameters in Temporal Fusion Transformer for Accurate Land Subsidence Prediction

This study looks at how varying hidden layer sizes affect the accuracy of the Temporal Fusion Transformer (TFT) model in forecasting land subsidence. Accurate projections are critical for mitigating the adverse effects of land subsidence on infrastructure and environmental management. This study tested hidden layer sizes of 8, 16, and 32 units using datasets from the Bandung and Yogyakarta regions. The model was trained using the Adam optimization algorithm with early stopping and learning rate scheduling to ensure convergence and prevent overfitting. Model performance for each configuration was evaluated using key metrics, including Mean Squared Error (MSE), Root Mean Squared Error (RMSE), Mean Absolute Error (MAE), and R-squared (R2). The results show that increasing the hidden layer size significantly improves model accuracy. Specifically, a hidden layer size of 8 yields an MSE of 0.427 and an R2 value of 0.283, while a hidden layer size of 32 yields much better performance, with an MSE of 0.034 and an R2 value of 0.872. In addition, the average land subsidence prediction values for the Bandung and Yogyakarta datasets show that a larger hidden layer size produces more reliable predictions, with a minor prediction error at a hidden layer size 32. In conclusion, this study highlights the importance of fine-tuning the hidden layer's size to improve the subsidence model's prediction accuracy. The findings provide valuable insights and practical guidance for geotechnical and environmental planners in making appropriate decisions regarding subsidence management.