Model Predictive Control when utilizing LSTM as dynamic models

The prediction model is the most important part of an MPC strategy. The accuracy of such a model influences the quality of predictions and control performance of the algorithm. In some practical cases, a model based on physical equations is not available, or is not easy to get all parameters, or its complexity could affect the real-time computation of the control signal. For this reason, the use of black-box models within a MPC framework becomes attractive, since to fit such models only input and output data are needed. Questions like: Is it possible to use LSTM's as predictors?, How to implement it?, What is the best way to compute derivatives?, Which solvers and tools are recommendable?, How to ensure the real-time capability?are discussed in this work.

Automated summary

The prediction model plays a vital role in MPC strategies as its accuracy directly impacts the quality of predictions and control performance. In cases where a model based on physical equations is not available or difficult to obtain all parameters, using black-box models within the MPC framework is an attractive alternative, as they only require input and output data. This paper discusses questions such as the feasibility of using LSTM as predictors, implementation methods, computation of derivatives, recommended solvers and tools, and ensuring real-time capability.