The reliability of power transformer has a significant impact on the reliability and security of the whole power grid. Existing research usually utilize the historical failure data and Weibull distribution to predict transformer's failure probability. However, history data-based Weibull distribution did not consider different service conditions of distinct transformers, resulting in inaccurate prediction results. In this paper, a cumulative modified Weibull model is proposed based on the environmental and operational data of the transformer. Parameters including ambient temperature, humidity and partial discharge are incorporated into the general Weibull model to better characterize the health condition of the transformer. During the operation, the corrections obtained by the modified model for each time sample are accumulated. By combining the history data-based failure prediction with the monitoring data-based correction, we can improve the accuracy of transformer reliability analysis. The cumulative modified model is verified by the monitoring data in real substations.