The invention discloses a fault diagnosis method for a power grid based on an improved Bayesian Petri net, and the method comprises the steps: firstly introducing a time sequence correlation rule, andstarting from the quantification angle to describe the time sequence relation between all places and a logic rule; secondly carrying out the reasoning of the time sequence constraint correlation of alarm information through a protection action rule under the condition that fault information is not complete, and giving an algorithm for recognizing the loss of alarm information, a breaking point and false alarm; finally achieving the positioning of a fault element quickly and accurately through combining with a Bayesian network. According to the invention, the method makes the most of the information in an alarm signal, and a fault diagnosis model which is based on the improved Bayesian Petri net and gives consideration to the time sequence correlation rule is proposed on the basis of the deep analysis of the time sequence attribute in the alarm information. The method reduces the uncertainty of an alarm result, is strong in resistance to interference, and can greatly improve the accuracy of a diagnosis result.