A Spatiotemporal Graph Convolutional Network for Predictive Control With Stability Guarantees

Abstract

Since controlled systems are becoming large-scale, complicated, and highly coupled, spatiotemporal features are becoming increasingly essential for accurately representing the system’s dynamics. Recent years have witnessed numerous attempts to enhance predictive models with spatiotemporal features. Inspired by the powerful spatial feature extraction capabilities of graph neural networks, this brief proposes a model predictive control (MPC) method based on graph convolutional networks (GCNs). Specifically, a novel GCN-based predictive model is first proposed, which can extract both spatial and temporal features. Notably, the model can also adaptively learn the spatial graph structure from time series data without requiring any prior knowledge. Then, combined with the MPC framework, the gradient descent (GD) method is introduced to handle the optimization problem. Moreover, stability and feasibility analyses have been conducted to guarantee the effectiveness of the proposed method during practical applications. The experimental results indicate the strength and reliability of the proposed method, which can greatly improve control accuracy compared to other learning-based MPC methods.