Feasibility of data-driven and mechanistic models for control studies in the Hias Process

This article assesses the feasibility of four data-driven and mechanistic models for developing control strategies for a biological nutrient removal process, the Hias Process. The control objective is to optimize aeration, the main energy cost of the process, using cascade loops with PID controllers. The main modeling challenges are the nonlinearity of the process, and the diurnal pattern of the water flowrate and nutrients, which makes the process time-variant. Dynamic time-lagging was applied to the real data to compensate for the time-delays. Based on our prior research, the most promising modeling methods were selected: support vector machine regression (SVM), Gaussian process regression (GPR), transfer functions (TF), and reduced order activated sludge models (rASM). These were trained and validated using measurements of dissolved oxygen and chemical oxygen demand at intermediate process stages. TF models fit the test data well, while SVM and GPR models suffered overfitting. The rASM model requires further refinement. Step response analysis around the nominal operation point demonstrated that SVM and GPR models exhibit faster dynamics and different response amplitudes compared to the TF and rASM models. Further work is suggested in development and testing of control strategies with the TF model.