Predição da condutividade da água purificada para uso na indústria farmacêutica

Abstract

Purified water plays a critical role in the pharmaceutical industry, being extensively used in both drug manufacturing and equipment cleaning. Ensuring the quality of this resource is essential to prevent contamination and maintain the efficiency of production processes. This study developed a predictive model for water conductivity, one of the key quality parameters, using Long Short-Term Memory (LSTM) networks. The study was based on data collected from a Water Treatment Station (ETA) at NUPLAM. The process involved data collection and preprocessing of time series data, including outlier removal and handling of missing data, followed by modeling and result evaluation. Three experiments were conducted: in Experiment 1, different input feature configurations and time lags were tested, achieving an MAE of 0.055 ± 0.002 and RMSE of 0.095 ± 0.001 for a 4-hour prediction horizon; in Experiment 2, hyperparameter optimization was evaluated but did not result in significant improvements; and in Experiment 3, focusing on different prediction horizons (1 to 6 hours), the best results were obtained for a 1-hour horizon, with an MAE of 0.026 ± 0.001 and RMSE of 0.049 ± 0.001. These results indicate that the model is more effective for short-term predictions. This study contributes to improving monitoring and control processes in pharmaceutical industries, including NUPLAM, and academically enhances the study of time series prediction for water treatment systems.