Traditional methods for construction progress prediction in smart construction sites often lack accuracy and adaptability. This study proposes a hybrid prediction model to enhance precision and robustness in complex environments.
An adaptive multi-population particle swarm optimization–long short-term memory (AMPSO-LSTM) model was developed by integrating the adaptive multi-population particle swarm optimization (AMPSO) algorithm with a long short-term memory (LSTM) network. Key site factors, including environmental conditions and resource allocation, were analyzed as model inputs, and AMPSO was used to optimize key LSTM hyperparameters, including the learning rate and the number of hidden neurons. The model was validated on tunnel construction data and compared with benchmark and reference models including LSTM, PSO-LSTM, BO-LSTM, BP, AMPSO-BP, ELM, GA-BP, RF, GRU, and an Informer-like baseline.
For the present tunnel case dataset, the AMPSO-LSTM model achieved an RMSE of 0.11061 and an R2 of 0.8337, while the monthly prediction error was 2.8%. Comparative results indicate that the proposed framework performs best among the selected benchmark models in terms of RMSE and R2, whereas the GRU model achieves the lowest MAE and the RF model yields the smallest absolute bias in MBE. These results suggest that the main advantage of AMPSO-LSTM in the present case lies in overall fitting accuracy and error reduction, rather than uniform superiority across every evaluation metric.
This study integrates adaptive swarm-based hyperparameter tuning with LSTM sequence modeling in a unified forecasting framework for tunnel construction progress prediction. The proposed AMPSO-LSTM model provides a practical case-based solution for progress estimation and schedule-related support under limited and dynamic engineering data conditions.
