As a new type of ecological technique in wastewater treatmen t field, Constructed Wetland was gradually put into application and received growing attention. Since the wastewater purification process in constructed wetland was a complex and nonlinear state affected by many interactional factors, it was hard to establish exact mathematics model to carry out multi-factor analysis and determine best operation condition by traditional means. Based on plenty of reliable experimental data, genetic neural network was first tentatively utilized to simulate the pollutant removal system of wetland, and some key problems were discussed, such as how to determine optimally the topological structure, sample scale and unitary method for training data, etc. Optimized GA-BP network was established to simulate orthogonal test of wetland system. According to the results of orthogonal test, best operation condition was decided and the factors(e. g.water level, hydraulic retaining time, etc)were classified. Therefore, proper feasible methods of pollutant removal were put forward.