Long-Distance Shield Tunnelling Performance Prediction Based on Informer

Shield performance prediction plays a critical role in construction decision-making. However, current models suffer from significant performance degradation in long-distance prediction. To address this gap, we propose a novel Long-Distance Shield Performance Prediction model (LSPP), which leverages the long-term prediction capabilities of Informer. The LSPP model incorporates conventional monitoring data, tunnelling parameters, and stratigraphic spatial information and is optimized using a ProbSparse self-attention mechanism and dynamic decoding techniques. A series of experiments demonstrate that LSPP significantly outperforms traditional models, such as LSTM and GRUs, particularly in long-distance predictions and under conditions of stratigraphic changes. Notably, the model achieves an R2 of 0.82 when predicting penetration after six rings, making it highly accurate and stable for engineering decision-making.