Temporal Convolutional Network-Based Spatiotemporal Modeling Method for Distributed Parameter Systems
提出一种结合时间卷积网络和长短期记忆网络的深度学习方法，用于分布参数系统的时空建模，在热过程和杆反应过程实验中优于已有方法。
Spatiotemporal modeling for distributed parameter systems (DPSs) described by partial differential equations (PDEs) is challenging due to infinite dimensional nature and strongly nonlinear spatiotemporal dynamics. In this study, a novel spatiotemporal modeling method based on deep learning network is proposed for DPSs. The deep learning network is composed of stacked temporal convolutional networks (TCNs) and long short-term memory (LSTM), where TCNs are developed to enhance the spatiotemporal feature representations and LSTM is used for prediction. The structural knowledge of the first-principles model is incorporated in the proposed deep learning network. The learning of the spatiotemporal feature no longer requires separate process. A moving window method is used to depict the dynamics of the most recent process, where the time-dependent correlation can be identified. The experimental results conducted on the thermal process and rod reaction process demonstrate that the presented approach outperformed those published methods in terms of root relative squared mean error (RSME), indicating its super performance.