In order to improve the forecasting precision of electric load data and the precision of characteristic analysis， dirty electric load data must be identified and justified early. A model for dirty data identification and justification is put forward based on the modified ART2 network. Pattern drifting phenomenon of the traditional ART2 neural network is improved in this model according to the idea that different distances between the sampling data and the clustering center of the same class exert different influences on the excursion of the clustering center. An identification subsystem is added to the ART2 model according to the residual error theory and inherent characteristics of the electric load curve. In the model proposed， electric loads are classified and the feature curves are picked up using the traditional ART2 neural network. Then the identification subsystem is used to identify and justify the dirty electric load data. Test results of actual data demonstrate the validity and feasibility of the modified ART2 network.