Abstract: Construction of reservoirs across river basins has profoundly changed the temporal and spatial distribution of river runoff. The operator of a downstream reservoir cannot obtain in real time the discharge plan of the upstream ones, as reservoirs are often operated by different owners. This not only makes it difficult to prepare an operation plan for the downstream reservoir, but also imposes a significant impact on its safety. This paper describes an adaptive-moment-estimation-improved deep neural network (Adam-DNN) simulation method for extracting operation rules from the historical operation data of different reservoirs and simulating their operation process by these rules. We verify this new method through comparison of its calculation results with those of the back-propagation (BP) neural network. Case studies reveal its average relative errors of 8%, 11% and 10% in discharge simulations for the reservoirs of Guanyinyan, Jinpingyiji and Ertan, respectively, much lower than the BP’s counterparts. The results show our Adam-DNN method can provide a new way to explore the operation rule of a reservoir with unknown operation plan.

Key words: runoff change, adaptive moment estimation, improved deep neural network, multi-reservoir simulation, reservoir operation rules