Modelagem preditiva multi-saída de um reator contínuo de tanque agitado

Abstract

This work presents the development and evaluation of a multi-output predictive model based on a Long Short-Term Memory (LSTM) neural network to represent the dynamics of a Continuous Stirred-Tank Reactor (CSTR). The study uses the CSTR dataset from the DaISy database, composed of 7500 samples with a sampling interval of 0.1 minute, equivalent to 6 seconds. The proposal aims to simultaneously estimate the process concentration and temperature from the recent operating history and is analyzed as a data-driven predictive component. The implementation was developed in Python, using cooling flow rate, concentration and temperature as historical inputs, organized into temporal windows of 50 samples. The data were standardized using mean and standard deviation computed only from the training segment, and the model was trained with a mean squared error loss function, Adam optimizer, early stopping, learning rate reduction and best model checkpointing. The evaluation considered one-step-ahead tests, external testing, recursive evaluation and multi-step forecasting, always comparing the LSTM network with a persistence model. The results indicated that the LSTM network was better able to estimate both output variables, following the process dynamics more closely than the persistence strategy, especially for horizons longer than one step. It is concluded that the proposed approach is suitable as a predictive component for the analyzed CSTR, within the limits of the dataset used. However, the developed model does not constitute a complete digital twin by itself, since it does not include real-time integration with a physical plant, continuous operational monitoring or automatic actuation on the process.