Model-based recurrent neural network for modeling nonlinear dynamic systems.

Area of Science:

• Control Systems Engineering
• Machine Learning
• Artificial Intelligence

Background:

• Dynamic systems modeling is crucial in various engineering fields.
• Traditional recurrent neural networks (RNNs) can be computationally intensive for training.
• Incorporating prior analytical knowledge into neural networks remains a challenge.

Purpose of the Study:

• To introduce a novel Model-Based Recurrent Neural Network (MBRNN) for dynamic system modeling.
• To leverage analytical plant knowledge within a neural network architecture.
• To improve training efficiency compared to standard RNNs.

Main Methods:

• Developed a fixed-topology MBRNN based on the linearized state-space model of a plant.
• Utilized Gaussian radial basis functions (RBFs) for activation functions to represent nonlinearities.
• Trained the MBRNN by adjusting RBF weights to modify activation function contours.

Main Results:

• The MBRNN successfully modeled dynamic systems in demonstrated examples.
• MBRNN achieved significantly shorter training times than ordinary recurrent networks.
• The fixed topology, independent of training, contributed to training efficiency.

Conclusions:

• MBRNN offers an efficient and effective approach for dynamic system modeling.
• Integrating analytical knowledge into neural networks can enhance performance and reduce training demands.
• The MBRNN architecture provides a promising direction for data-driven modeling with prior system insights.