Leveraging Deep Learning and Rolling Simulation for Adaptive Reservoir Operation Under Uncertainty

Abstract

Reservoir operation rules play a crucial role in reservoirs’ scientific management and safe operation, which can provide reliable scheduling guidance information for decision-makers. Existing reservoir operation rules derived from implicit stochastic optimization methods often fail to account for the cumulative impact of decision errors on the reservoir, rendering them less effective in actual reservoir operation. In this study, a novel rolling simulation strategy with decision error correction capability is proposed. Furthermore, the strategy is combined with a shared weight long short-term memory network, a novel deep learning method for data mining, to extract reservoir operation rules. The proposed method is applied to a real-world case at the Guangzhou Reservoir in the Beipanjiang River Basin. Experimental results demonstrate that the proposed method outperforms comparative model in terms of power generation efficiency and water level deviation indicators. Moreover, the rolling simulation strategy has been proven to effectively reduce cumulative decision errors in reservoir operation rules, while the shared-weight long short-term memory network has demonstrated excellent data-mining capabilities. In conclusion, the method proposed in this paper can provide more reliable operation rules and decision support for decision makers, and realize the improvement of power generation efficiency of hydropower system.

Keywords

• Reservoir operation
• rule derivation
• rolling simulation strategy
• deep learning
• Guangzhou reservoir