Bayesian-optimized Neural Networks and their application to model gas-to-liquid plants

Neural networks are robust modeling techniques widely used in regression and classification applications. Model selection criteria for neural networks have typically been applied towards the optimization of the network's weights and biases, and not for the structure of the network itself. In this work, a systematic framework is proposed for the estimation of an optimal neural network structure using Bayesian optimization, denoted as the Bayesian-optimized Neural Network (BONN). In addition, unlike typical applications in literature, the purpose of the BONN framework is to provide a non-linear alternative to Principal Component Analysis (PCA), which is a modeling technique used primarily for dimensionality reduction and noise filtration. These operations are necessary to build a robust model capable of accurately capturing the true dynamics of a complex multivariate system in the presence of noise. As a result, two case studies have been carried out to benchmark the performance of the BONN framework in comparison to PCA. The first case study, applied to non-linear synthetic data, showed that the BONN model was able to approximate the noise-free behavior of the data more accurately than PCA and at a smaller dimensionality. The second case study, applied to real data gathered from a bench-scale Fischer–Tropsch process, showed that BONN was visually more capable of predicting the true behavior of process variables in the presence of noise without overfitting.