Abstract: Due to the problems such as strong continuity and high complexity of the data generated by the thermal power control process,and the difficulty of establishing a semantic correlation between the model behavior of recurrent neural networks and the actual control process,the model debugging,optimization and semantic analysis cannot be carried out intuitively.Visual analytics in recurrent neural networks modeling for system identification is applied and a visual analysis system called iaRNN is proposed.Firstly,by visu-alizing the activation value distribution and coverage of hidden units,the combined view for model evalua-tion is designed to support the evaluation of model performance in many aspects including inside and out-side.Then,visual views are designed from the perspectives of temporal evolution,sensitivity analysis,etc,to support the exploration of model response behavior to thermal power control parameters.Finally,based on the symbolic representation of time sequences and clustering analysis,a method for exploring association patterns between strong time-dependent real-valued time series and hidden units is proposed.A case study using real power plant data is conducted to verify the effectiveness of iaRNN in assisting users to understand the working mechanism of the model and diagnose model defects.