Towards Data-Driven Modelling of Saulekilen Wastewater Treatment Plant based on Artificial Intelligence

Abstract

The processes at a wastewater treatment plant (WWTP) are complex systems thatclean the wastewater before it is released into the environment. Total phosphorous(TP), biological oxygen demand (BOD) and chemical oxygen demand (COD) arecritical measurements of the water quality. Measuring BOD and COD are too slowto effectively control the wastewater treatment processes (5 days and a couple ofhours respectively) as the wastewater treatment process at Saulekilen WWTP usesabout 30-40 minutes from influent to effluent. Measurements of effluent TP are alsoperformed and are quicker but controlling by the use of effluent TP is too slow. Henceit is a goal for this thesis to propose models that can predict the values of BOD, CODand TP quicker by the use of multiple linear regression (MLR) and artificial neuralnetworks (ANN). Results of this thesis indicate that measurements of effluent TPcan be predicted with reasonable accuracy. Further, the influent BOD and CODmeasurements can be predicted with good accuracy. These promising predictionsmay be important in further work for an improved wastewater treatment.A sensitivity analysis has been performed to determine which input parameters beingthe most effective in modelling the output. Input parameters in this thesis haveconsisted of TP, water discharge, the local temperature and the date for predictionsof TP, COD and BOD. Measurements of influent BOD and COD has also been usedas inputs for prediction of effluent BOD and COD, while total solids (TS), totalpercent solids (TS(%)), reject water, pH, the process additive PIX and turbidity hasbeen used as inputs for predictions of effluent TP. The results of the models havebeen considered with respects to correlation, mean squared error (MSE) and meanabsolute percentage error (MAPE).