Data-driven models to enhance phosphorus removal in an edible oil wastewater treatment plant.

Abstract

The discharge of effluents with high phosphorus concentrations into water bodies can lead to significant environmental problems. Addressing this challenge is critical, particularly in developing countries, where independent waste water treatment plants (WWTPs) are prevalent and often lack sensors for continuous monitoring, making their operation and control more difficult. Within this context, this thesis explores the use of data-driven models to enhance the operation of an edible oil WWT for phosphorus removal. In the absence of phosphorus online monitoring, a model that forecasts phosphorus concentration would enable the plant to anticipate when additional treatment, such as physico chemical removal, is required to meet the recommended phosphorus standards. For this purpose, various machine learning (ML) and deep learning (DL) techniques are evaluated, including Linear Regression (LR), Random Forest (RF), Support Vector Regression (SVR), and LongShort-Term Memory (LSTM) neural networks. Given the non linear nature of wastewater treatment processes, several feature selection methods besides Pearson correlation are explored, such as Spearman correlation, RF feature importance ranking, and causal inference for time series. The models are evaluated across different phosphorus concentration ranges, as errors in predicting high concentrations have a greater impact on plant operations. Results show that LSTM networks with selected features out perform other models for forecasting next-day phosphorus concentration, though challenges remaining accurately predicting peak concentrations. Additionally, surrogate optimization is used to estimate appropriate chemical dosages for the operation of the plant’s physic-chemical phosphorus removal (PPR) system. The surrogate model is built using simulated data from Bio Win, and data acquisition is facilitated by the developed API Bio2Py (BioWin to Python) that integrates BioWin with Python. The surrogate model, implemented using a Feed forward Neural Network (FNN), demonstrates good performance and is successfully integrated into an optimization tool that provides rapid chemical dosage estimations. However, the tool tends to overestimate aluminum sulfate dosages, indicating the surrogate model needs further improvement. The presented data-driven tools can enable faster decision-making, leading to more efficient and cost-effective operations. Furthermore, the presented approaches could be applied to other WWTPs to enhance their phosphorus removal processes, offering significant potential for broader applications in the field of wastewater treatment.