| Title |
Short-Term Dam Inflow Prediction Using Machine Learning with State-Space-Based Input Reconstruction: Case Study of Hapcheon and Namgang Dams in the Nakdong River Basin |
| Authors |
이동민(Lee, Dongmin);오랑치맥 솜야(Uranchimeg, Sumiya);권현한(Kwon, Hyun-Han) |
| DOI |
https://doi.org/10.12652/Ksce.2025.45.4.0469 |
| Keywords |
댐 유입량 예측; 기계학습; 상태-공간 모형; LSTM; SVM Dam inflow forecasting; Machine learning; State-space model; LSTM; SVM |
| Abstract |
This study develops a machine learning-based inflow forecasting model by reconstructing input time series using a state-space embedding method that captures nonlinear and nonstationary behaviors. The target sites are the Hapcheon Dam and Namgang Dam in the Nakdong River Basin, with areal rainfall, upstream water level, and inflow data used as model inputs. The delay time and embedding dimension were determined using the Average Mutual Information (AMI) and False Nearest Neighbor (FNN) methods. Long Short-Term Memory (LSTM) and Support Vector Machine (SVM) models were applied to predict inflows with lead times ranging from 1 to 6 hours. For the Hapcheon Dam, the SVM model achieved a 6-hour lead time prediction with a correlation coefficient(CC) of 0.887 and an RMSE of 88.9 m3/s, significantly outperforming LSTM (CC = 0.655, RMSE = 234.51 m3/s). For the Namgang Dam, the 6-hour prediction showed a smaller performance difference between SVM (CC = 0.887) and LSTM (CC = 0.858), but SVM tended to underestimate peak inflows. These results show that SVM performs well under normal to moderate inflow conditions, while LSTM is more effective at capturing extreme peaks. This study presents a practical inflow forecasting framework by combining optimized input reconstruction with machine learning, and future improvements are anticipated through integrating weather forecasts and spatial data. |