Author(s): Quan Le Huu Minh; Giang Nguyen Tien; Hanh Nguyen Duc; Huy Dao Ba; Anh Nguyen Thi Phuong; Chi Nguyen Que

Linked Author(s):

Keywords: Reservoir inflow forecasting; Machine learning; Uncertainty analysis; Multi-step ahead forecasting

Abstract: The Da River Basin is a transboundary river basin, covering an area of 53,000 km2, in which, 49% of its upstream area locates in China and the remainder locates in Vietnam. Numerous large reservoirs have been constructed in this basin, mainly for electricity generation. In Vietnam, there are five major reservoirs namely Lai Chau, Hoa Binh, Son La, Ban Chat, and Huoi Quang. Among these, Lai Chau and Ban Chat are the two uppermost reservoirs located on the main upstream branches of this interconnected reservoir system. Accurate inflow forecasts for these two upstream reservoirs are crucial for efficiently operating the entire system. Given the limited availability of observational data from the upstream region, forecasting approaches based on traditional statistical methods and, more recently, machine learning techniques are deemed appropriate. This study aims at developing and selecting forecasting models for inflow prediction to these two reservoirs. The models considered include several established machine learning algorithms namely Random Forest (RF), LightGBM (LGBM), Support Vector Regression (SVR), and Long Short-Term Memory (LSTM). The performance of the models was assessed through standard evaluation metrics (NSE, RMSE, MAE, KGE) complemented by an in-depth uncertainty analysis. The results indicate that for Lai Chau Reservoir, the LGBM model provided the most accurate prediction for one-day-ahead forecasts (NSE = 0.87, RMSE = 178, MAE = 108, and KGE = 0.92), the SVR model provided the most accurate prediction for 10 days one step ahead forecasts (NSE = 0.64, RMSE = 297, MAE = 196, and KGE = 0.80). Meanwhile, for Ban Chat Reservoir, the RF model achieved the best one-day-ahead prediction (NSE = 0.74, RMSE = 87, MAE = 34, and KGE = 0.70). Over a 10-day forecasting interval, the RF model continued to provide the most accurate results, with NSE = 0.68, RMSE = 56, MAE = 36, and KGE = 0.80. When applying these best-performing models (LGBM and RF for daily streamflow forecasting at Lai Chau and Ban Chat, respectively, and SVR and RF for 10-day average streamflow forecasting at Lai Chau and Ban Chat) for multi-step ahead forecasting, the results reveal a decline in performance as the lead time increases. Generally, forecasts become less reliable (NSE < 0.5, KGE < 0.5) when reaching the seventh forecasting step.

DOI: https://doi.org/10.64697/978-90-835589-7-4_41WC-P1981-cd

Year: 2025