Abstract: The potential transformer（PT） is the key device for voltage measurement in power system. Online evaluation of the measurement error of PT will help maintain the fairness and justice of electricity trade settlement. Among them, the data-driven online evaluation method has good engineering application prospects because of its superior evaluation performance. However, there are problems such as failure to consider whether the modeling data contains abnormal data and the inability to quantitatively evaluate. To this end, this paper proposes a quantitative evaluation method based on optimal modeling and deep belief network（DBN）. This method adopts unsupervised clustering technology to determine the ideal modeling data set, and utilizes the deep belief network for training to obtain the PT quantitative evaluation model, and then, analyzes the real-time output signal of PT to achieve error quantitative evaluation. Experiments show that this method can effectively detect the abnormal data in the modeling data, and accurately monitor the measurement error status of the 0.2-level PT, and realize the accurate evaluation of the measurement performance of in-service PT.