Distillation Column Temperature Prediction Based on Machine-Learning Model Using Wavelet Transform

This study presents a machine-learning-based prediction model for distillation process operation data using wavelet transform. The process operation data collected from a distillation column contain noise due to sensor errors. Developing a machine-learning model using noisy data reduces the accuracy of the model; therefore, the data should be denoised. Denoising was achieved using wavelet transform, and a long short-term memory (LSTM) machine-learning model was developed. Wavelet transforms generally decompose data into high- and low-frequency components using wavelet functions with various frequencies. The high-frequency components are the details comprising noisy data, and the low-frequency components correspond to the approximations of the original data. The approximations were used to develop the LSTM model. Depending on the type of wavelet function used for decomposition, the denoised values varied and affected the model accuracy. Case studies were conducted using various wavelet functions to develop models with optimum prediction performances. By applying the optimal wavelet transform to the LSTM model, the prediction performance improved by 10%.