Abstract: Transmission line icing accident is very destructive to the safe operation of power grid system. The prediction of icing thickness can effectively guide the ice-resistant work of power grid. In order to realize the accurate short-term prediction of icing, a combined prediction model of RF-APJA-MKRVM considering time cumulative effect is proposed from the point of view that line icing is a time accumulation process to predict different icing stages. Firstly, the random forest algorithm is used to select the most important factors affecting icing, and the adaptive parallel Jaya algorithm is used to optimize the multi-kernel relevance vector machine to establish the combined prediction model of icing growth rate. Finally, on the basis of the combined prediction model, by taking into account the time cumulative effect of icing growth and the initial thickness of different stages, the prediction results of final icing thickness are obtained. Based on the related data of actual icing collected by online monitoring system of Guizhou Power Grid, it is verified that the average root mean square error of the prediction model in the icing growth, stability and melting stages are 0.130, 0.121 and 0.137, respectively, which confirms the effectiveness of the prediction model. Compared with the similar algorithms, its accuracy has been greatly improved, and the icing prediction in different stages is distinguished, which can provide a certain reference for transmission line deicing work.