Abstract: Stochastic simulations of the errors in reservoir inflow process forecasts with multiple forecast periods become more difficult as the number of dimensions increases. To examine the variation trends of the errors accurately and quickly, we first generate the characteristics of complex high-dimensional data through numerical simulations using the neural network coupled with low-dimensional hidden variables and the Variational AutoEncoders (VAE) method. Then, we develop a stochastic simulation model of the forecast errors of reservoir inflow process based on VAE. This model is compared with the improved Gibbs method in a case study of the Jinping Ⅰ hydropower station. The results show that it gives better agreement of the mean, standard deviation, and variation coefficient with the real error sequence, and its computational time reduces by 69% to 94% compared with the improved Gibbs method. These results provide more information for hydropower station regulation considering uncertainty in reservoir inflow forecast.

Key words: reservoir inflow, forecast errors, multi-dimensional stochastic simulation, Variational AutoEncoders, improved Gibbs method, Jinping I hydropower station