Surrogate model of flooding loss to alleviate computational burden in reservoirs operation

Joint operation of reservoirs can effectively reduce flood loss. However, the traditional reservoir operation model considers downstream flood peak rather than flooding loss, due to the heavy computational burden of hydrodynamic simulation. To addressed this issue, the machine learning-based surrogate model, which can accelerate the hydrodynamic simulation, is used to reduce flooding loss by coupling with the reservoir operation model. The machine learning surrogate model can quickly simulate flooding loss, but leads to the reservoir operation model no longer meeting the Markov property. As a result, dynamic programming (DP) and its improved algorithms are unable to deal with this optimization problem. Thus, DP only generates an initial solution, which can be further refined by the pattern search algorithm to minimize flooding loss. The Centianhe and Shuangpai Reservoirs on Xiaoshui River Basin, Hunan Province, China were selected as the study area. Results showed that: (1) the surrogate model can shorten the flooding loss calculation time from the minute level of the hydrodynamic model to the millisecond level, while ensuring accuracy of average RMSE 0.629 m and the R² 0.83, and (2) the proposed reservoir operation model significantly reduces flooding loss. Compared with traditional models, the proposed model reduces flooding loss by 16.28 % and 13.74 % under the design floods of 3-year and 5-year return period, respectively. Even the proposed method can be improved in terms of model generalizability and accuracy, it provides a valuable model for high flood risk basins by shifting the reservoir operation objective from flood peak shaving to flooding loss reduction.