Activated sludge wastewater treatment is a highly complex process, and variations in wastewater flow rate and its composition, combined with time-varying reactions of microorganisms, make this process non-linear and dynamic. Due to this fact, process control is very sensitive but a proper model could present a solution for this problem. In this paper several models have been tested for prediction of COD in effluent of wastewater treatment plant "Velkom" in Velika Gorica. Proposed models are: Multivariate Linear Regression model (MLR) with 16 predictor variables included, Piecewise Linear Regression model (PLR) with 17 predictor variables and RBF Neural Network model (RBF NN) with 18 predictor variables included. Analysis of results show most successful prediction of process behaviour by RBF NN, with correlation 0,83, which is then used for estimating optimal operative conditions (wastewater flow rate should be in range 120-145 L/s and concentration of activated sludge dry matter between 2,5 – 4 kg/m³).