No abstract

No abstract

This paper addresses the problem of determining achievable set‐points for artificial neural network (ANN)‐based model predictive control (MPC) designs. In particular, this work considers a case where a first‐principles model may not be readily available for a nonlinear process, while sufficient closed‐loop data containing possibly correlated outputs is available, such that an ANN‐based model that captures the nonlinear dynamics reasonably well can be identified. The paper addresses implementation aspects with such an ANN‐based MPC design—specifically that of ensuring that achievable set‐points are prescribed to the MPC. The key idea is to perform principal component analysis (PCA) on the training data in order to recognize existing collinearity and determine the upper confidence limit of squared prediction error (SPE) statistic. An optimization problem subject to the SPE constraint is then defined to calculate the achievable set‐points, that can in turn be provided to an MPC design. The efficacy of the proposed approach is illustrated via implementations on a chemical reactor example. The results reveal the superior tracking performance of MPC using the achievable set‐points over the case where arbitrarily prescribed set‐points are used in the MPC implementations.

This work addresses the problem of implementing model predictive control (MPC) in situations where the training data available for modeling contains possible correlations, and an artificial neural network (ANN)-based model is being used. In particular, we consider a problem where data sets are collected from a process that operates under the closed-loop condition in which correlations are induced between several input and output variables. In this situation, if the correlation problem is not addressed, manipulated inputs (calculated by MPC without considering the specific correlation in the input space) and independently prescribed set-points may require predictions in regions where the model is not trained, resulting in a poor closed-loop performance. To address this issue, principal component analysis (PCA)-based strategies are applied to both the input and output spaces in a way that maintains model validity. To that end, a new constraint on the squared prediction error (SPE) is incorporated into the ANN-based MPC optimization problem to make control actions follow the PCA model built using the training input data. Next, a PCA model is developed using the training output data, and then an optimization problem subject to the SPE constraint is defined to calculate set-points which are achievable. The effectiveness of the proposed ANN-based MPC to track these set-points is demonstrated using a chemical reactor example. Finally, a new autoencoder-based strategy is proposed to compute the achievable set-points. This is performed by replacing the PCA-based constraint with the autoencoder-based constraint in the optimization problem to calculate the set-points. The results indicate that the ANN-based MPC performance is improved when the autoencoder-based set-points are used.

This paper addresses the problem of system identification for heating, ventilation, and air conditioning (HVAC) systems using a relatively small amount of data for the zone under consideration, by leveraging larger datasets for similar zones. To this end, a hybrid machine learning approach is developed where a pre‐trained recurrent neural network (RNN) model, trained on a large amount of data from a representative zone, is leveraged to build models for the other zones using a smaller amount of data. This is achieved by developing a hybrid model that integrates the pre‐trained RNN model with the models built using the subspace identification (SubID) technique to predict the residuals (differences between the real outputs and the predicted outputs from the pre‐trained RNN model) in the other zones. The effectiveness of the proposed hybrid approach is shown using real data collected from a multi‐zone fitness centre. The results demonstrate the superior performance of the hybrid approach over the cases where individual RNN and SubID models are directly developed using only the data from the zones in question.