Extracting operation behaviors of cascade reservoirs using physics-guided long-short term memory networks

Abstract

Study region: Qingjiang cascade reservoir, China. Study focus: Reservoirs regulate the natural streamflow to utilize water resources comprehensively. How to mine the existing massive reservoir operation data to describe human operation behaviors is a challenge. To address this issue, a data-driven method, Long short-term memory (LSTM), was used to simulate the reservoir outflow by inputting historical information. The physics-guided LSTM model, shortly named PG-LSTM, was formulated by using synthetic flood samples and physical constraints of water balance, boundary, and monotonicity. New hydrological insights: (1) PG-LSTM can reproduce historical outflow with seasonal variations, or predict outflow without lags, (2) knowledge of reservoir operations can guide LSTM with the reduction of negative flow occurrence and the accurate identification of operation behaviors under extreme hydrological conditions, (3) specifically, compared with conventional LSTM, gradient boosting regression tree and conventional reservoir operation, PG-LSTM can improve the Nash-Sutcliffe efficiency of cascade reservoir during the test period from 0.50, 0.20, and 0.17 to 0.54 in the reproduction scenario, and from 0.84, 0.26, and 0.17 to 0.85 in the prediction scenario with five-fold cross-validation method. The PG-LSTM is helpful to describe human operation behaviors of reservoirs.

Keywords

• LSTM
• Data-driven
• Extracting human behaviors
• Physics-guided