Mine water hazards threaten safety; burst-point inflow can spike rapidly. Inflow series are nonlinear, uncertain, and often small-sample. Shi et al. (2024), Scientific Reports, propose LSTM-Transformer. Summary below.
Scenario and goals
- Data: 245 records (2012–2022) from Baotailong mine, Heilongjiang (as in the paper).
- Goal: rolling inflow forecasts for warning and response.
- Challenge: overfitting; balance generalization and accuracy.
Method highlights
1. Preprocessing
ADF stationarity, EMA, Ljung-Box tests; min–max scaling; train/test ratios including 7:3 where LSTM-Transformer performed best in the study.
2. LSTM-Transformer
LSTM for long dependencies; Transformer self-attention for global structure; dense head for next-step prediction.
3. Hyperparameters
Random search + Bayesian optimization with L2 regularization and Adam.
4. Baselines
Compared with CNN, LSTM, Transformer, CNN–LSTM; the paper reports LSTM-Transformer leading on RMSE, MAE, and R² (e.g. ~18.2% / 21.3% MAE/RMSE reduction vs CNN—see original tables).
Practical takeaways
- Treat split ratio as a first-class experiment axis on small samples.
- Fusion can combine local dynamics and global patterns.
- Wire forecasts into flood warning workflows.
- Authors note limited dataset scale for generalization.
Reference
Shi, J.; Wang, S.; Qu, P.; Shao, J. Time Series Prediction Model Using LSTM-Transformer Neural Network for Mine Water Inflow. Scientific Reports 2024, 14, 18284.