The invention provides a power grid reliability sequential simulation method based on conditional kernel density estimation. The method adopts a maximal load capacity, a system state average duration time and a sound load percentage as MDP state parameters of system states. One time of traditional sequential simulation is performed under the condition of an annual maximal load level of the system, the MDP state parameters of all system states are used as MDP state parameter samples, and conditional probability density distribution is established. When the system load level or a load variation rule changes, the system states are extracted according to the conditional probability density distribution of the original MDP state parameters, computation is implemented by using the MDP state parameters under each system state, and thus the loss of load and the duration time of each system state are obtained, and reliable indexes are formed. According to the method, the system states are extracted according to the conditional distribution of the MDP parameters, and only the sizes of the maximal load capacity and the load under each state need to be compared, with no need to implement optimal power flow computation on each system state, and thus the sequential simulation time is greatly reduced.