The invention discloses a regional meteorological condition similarity-based large power network load prediction method. The method comprises the steps of firstly obtaining load prediction values of regions in a to-be-predicted day and a load and meteorological history data of a sample space recently; aggregating the regions with similar meteorological conditions according to a probability distance-based synchronous back-substitution elimination technology, dividing an aggregated region into q sub-regions with relatively large meteorological condition difference, then calculating an average proportional coefficient of the q sub-regions at a t moment point in a historical data sample space, and predicting proportional coefficients of the sub-regions at the same t moment point in the to-be-predicted day; and predicting a whole network system load at the t moment point by using the q sub-regions, building an optimal comprehensive model at the t moment point for q different prediction values, performing solving to obtain a final prediction result of the whole network system load at the t moment point, and building optimal comprehensive models for T moment points in the whole to-be-predicted day, thereby obtaining a whole day load prediction sequence. According to the method, the short-term load prediction accuracy of a power system can be improved.