Finding ways of optimizing coagulant dosage, for a more sustainable wastewater treatment process

Abstract: The wastewater treatment process at Oatly consists of several treatment steps of which, two are chemical treatment steps involving coagulation followed by flotation to remove phosphorous and COD. The coagulant dosage is mainly determined based on the operator’s experience and supported using results from jar tests. Jar tests are, however, highly dependent on the wastewater quality parameters making it difficult to apply the result if there are large variations in the composition of the wastewater, which is the case at Oatly. As wastewater parameters such as concentration of COD and Tot-P in the primary influent, pH, temperature, and flow rate all varies multiple jar tests have to be conducted daily and the coagulant dosage is often adjusted. Despite this, the BOD7 concentration of the final effluent is well within the concentration limits of the environmental permit. This in combination with the large amounts of coagulant used in the process has raised the question, is it possible to decrease the dosage and still fulfill the requirements of the permit? The large variations of the primary influent make this difficult to achieve as there is no simple linear relationship between the amount of coagulant used and the coagulation efficiency, instead, it is influenced by several parameters. Instead, machine learning is used where a random forest regressor algorithm is used to predict the concentration of COD in the effluent based on the flow rate, pH, temperature, PAC dosage, the influent concentration of COD, and Tot-P. To train the model data from the first chemical treatment step is used as more data is available compared to the second chemical treatment step. After training the model it is used to predict the COD concentrations of the effluent from the first flotation. The results show an r2 value of 0.85 and an RMSE value of approximately 49, indicating a strong relationship between the independent and dependent variables used in the model. The RMSE value on the other hand should preferably have been lowered, it means that the difference between the concentration predicted by the model and the real concentrations is 49 mg/L on average. To illustrate the possible benefits of the model it is used to estimate the COD concentration based on test data where the dosage of PAC is varied. The predicted COD concentrations are then compared to the COD concentration of the effluent with a 57 % removal efficiency. A 57% removal efficiency was used to determine what would have been a sufficient amount of COD in the effluent as this is the removal efficiency that was used when the WWTP was designed. The model predictions showed that it, according to the model, would have been possible to decrease the coagulant dosage and save around 280 kg of coagulant/day on average. The carbon footprint of PAC is approximately 0.54 kg CO2-eq/kg PAC so this would mean a decrease in CO2 emission by around 150 kg CO2-eq per day. Decreasing the dosage would also make it possible to save approximately 1000 SEK. The main suggestion for areas of future work is to increase the size of the dataset to improve the performance of the model. However, the model shows potential as a tool to predict the COD concentrations of the effluent and could be used to estimate the effect of different coagulant dosages, thereby opening up the possibility of optimizing the coagulant dosage.