Reinforced recurrent neural networks for multi-step-ahead flood forecasts

摘要: Summary Considering true values cannot be available at every time step in an online learning algorithm for multi-step-ahead (MSA) forecasts, a MSA reinforced real-time recurrent neural networks (R-RTRL NN) is proposed. The main merit of the proposed method to repeatedly adjust model parameters with current information including latest observed and model’s outputs enhance reliability forecast accuracy method. sequential formulation R-RTRL NN derived. To demonstrate its effectiveness, implemented make 2-, 4- 6-step-ahead forecasts famous benchmark chaotic series reservoir flood inflow North Taiwan. For comparison purpose, three comparative (two dynamic one static networks) were performed. Numerical experimental results indicate that not only achieves superior performance but significantly improves precision both case during typhoon events effective mitigation time-lag problem.