This paper introduces a novel Feature-Extended Neural Ordinary Differential Equation (Fe-NODE) method for modeling black-box control systems by incorporating control parameters into the ordinary differential equation. Since the input to the model should consider both the system state and the control parameters, this is a challenge to the modeling process. The primary contribution of the proposed method is employing feature expansion to add information to the model input and enhance the perception ability of the model while utilizing intermediate state supervision and teacher forcing to optimize the training process, and effectively identify potential models for black-box control systems. The experimental results demonstrate that the Fe-NODE method outperforms the baseline in prediction accuracy and modeling efficiency. Additionally, we investigate the impact of different feature expansion coefficients on the model’s prediction accuracy, identifying the optimal coefficient through comparative analysis. This approach provides a novel methodology for modeling black-box control systems and lays the foundation for model-based reinforcement learning to achieve optimal control.