LSTM-based prediction and parameter optimization analysis of sandstone subgrade compaction quality

Abstract

The study proposes an LSTM-based framework for next-pass Compaction Energy Value (CEV) prediction and subsequent energy planning during sandstone subgrade compaction. Field tests were conducted at two sites, each divided into 24 spatial areas. For each completed compaction pass, cumulative vibration input energy, moisture content, and measured CEV were used to characterize the historical compaction state. At Site 1, information from the first two completed passes was used to predict the CEV after the third pass, whereas at Site 2, information from the first three passes was used to predict the CEV after the fourth pass. The LSTM model achieved RMSE, MAE, and MAPE values of 1.008, 0.801, and 4.74%, respectively, for Site 1, and 1.292, 1.218, and 6.92% for Site 2. To evaluate model selection under the limited field dataset, LSTM was further compared with XGBoost and SVR using identical six-fold spatial cross-validation partitions. LSTM achieved the lowest out-of-fold RMSE, MAE, and MAPE, with values of 0.486, 0.356, and 2.17%, respectively. For subsequent energy optimization at Site 2, area-specific relationships between CEV and energy relationships were first used to estimate initial candidate energies, which were then iteratively adjusted through repeated forward prediction using the fixed LSTM model. After iteration, the proportion of areas with predicted CEV values within the prescribed control interval increased from 75 to 100%, while the predicted CEV range and standard deviation decreased by 75.9% and 82.3%, respectively. The resulting remaining energy demands were further converted into rolling plans, with one weak-vibration pass assigned to 7 areas and one strong-vibration pass to the remaining 17 areas. The proposed framework provides a quantitative link between next-pass CEV prediction, area-specific energy demand, and subsequent rolling planning for sandstone subgrade compaction.