Data-predictive Control of Multi-Timescale Nonlinear Processes

A novel big data-predictive control approach for nonlinear multi-timescale processes is presented in this paper. Multiple Dynamical Latent Variable Autoencoders (DLVAEs) are employed to approximate multi-timescale dynamics, utilizing timescale-based low-pass filtering and resampling of historical input-output data. The encoder in each DLVAE projects the nonlinear physical variable space onto a linear latent variable space, represented by a kernel space in behavioral system theory. During training, we not only impose kernel spaces and reconstruct data but also establish connections among latent variables from different DLVAEs at matching time-steps. Collectively, these multi-level latent variables span a wide prediction time horizon with limited (non-uniformly spaced) steps encompassing the current, near, and distant future. In online tracking control, we guide the latent variables from each DLVAE to their respective setpoints (derived from physical variable setpoints) while maintaining consistent physical variable values at matching time-steps, all within a linear framework.