The invention provides a power grid robust state prediction method based on multi-dimensional state matrix sliding matching. The power grid robust state prediction method comprises the steps of obtaining a node admittance matrix; acquiring SCADA instrument quantity measurement and PMU device quantity measurement, and constructing a state quantity set based on mixed state estimation; forming a historical state quantity database; predicting the future state of the nodes of the smart power grid by using a multi-dimensional state matrix sliding matching method to obtain a prediction result; adopting an improved power balance equation detection method to detect whether abnormal data exists in a current system or not, if yes, a residual distribution deviation degree detection method is adopted to judge whether the abnormality is caused by abrupt change of quantity measurement, if yes, correction is carried out, bad data are removed, and reliable data support is provided for state predictionin future. The method has better prediction robustness, information in a historical database can be effectively utilized, and the safety and reliability of the intelligent power grid and the capability of resisting malignant data and bad data are improved.