Operational Interval Extraction Based on Long-Short Term Memory Networks for Building More Feasible Reservoir Operation Models

Abstract

Advances in data analytics create an opportunity to enhance reservoir operation. A challenge arising is how to utilize operational data to form realistic constraints of the reservoir operation practice. To address this issue, a novel approach is proposed to extract operational intervals of reservoir outflow by a deep learning method, namely the physics-guided long-short term memory network. The knowledge-informed reservoir operation (KIRO) model was built by adding derived operational intervals of outflow as constraints for the traditional reservoir operation (TRO) model. KIRO couples (a) an optimization model to search for optimal operation schemes, and (b) operational intervals of reservoir operators' decisions based on realistic factors. China's Qingjiang cascade reservoir including Shuibuya, Geheyan, and Gaobazhou reservoirs is used as a case study. Results show that KIRO can yield more physically feasible operation schemes than TRO due to its additional constraints. Specifically, KIRO avoids excessive reservoir water level fluctuations and outflow variations compared with TRO. Moreover, the extracted operational interval can help uncover implicit demands of real-world operation, for example, the KIRO model accurately identified the cascade reservoir unit maintenance events from 31 January 2019, to 31 March 2019, and the operation schemes were aligned more closely with the power demands. This study provides a new method for building more feasible reservoir operation models based on deep learning.