Physics-informed neural network based modeling of an industrial wastewater treatment unit

Abstract

Wastewater treatment units consist of biological treatment with activated sludge and are subject to many disturbances such as influent flowrate, pollutant load and weather conditions bringing about many challenges for the modeling of such plants. Data-driven models may respond to these challenges at the cost of issues such as overfitting or poor fitting due to the lack of high-quality data. To benefit from the available physics-based knowledge and to eliminate the drawbacks of suboptimal and poor training, physics informed neural networks might be quite promising. In this work, artificial, recurrent and physics-informed neural network models are utilized for the wastewater plant in Tüpraş İzmit Refinery. For recurrent models with selected features based on correlation technique, test mean squared error is up to 82% smaller compared to the standard artificial neural network models. Physics-informed trained neural network models with selected features improved the test performance by decreasing mean squared error up to 87% with acceptable decreases in training performance which addresses its strength compared to fully data-driven models.