A prediction approach of waste-water concentration was proposed, in order to reduce costs or decrease monitoring devices, by minimum relative error performance index. Firstly, non-linearly data on low dimensional space was mapped into high dimensional space, then established the linear prediction model on the high dimensional space, finally an application example was provided to illustrate the method, and comparisons was made with the traditional least squares method or the current popular neural network method. Simulation shows the method is simple and effective.