Abstract: The smart grid big data environment provides an opportunity to solve the problems of short-term load forecasting model performance degradation and accuracy degradation over time. In light of this, a novel dynamic modeling methodology combining on-line SVR（OSVR） with Fisher information（FI）-based meteorological factors introduction and features selection（FS） was developed, and constructs a smart grid weather sensitive load forecasting model when the relationship between process variables changes rapidly. First of all, a concise OSVR learning algorithm was derived using Karush-Kuhn-Tucker（KKT） conditions in SVR. The OSVR algorithm efficiently updated a trained SVR function whenever a sample was added to or removed from the training set without retraining the entire training data. Secondly, a novel method based on Fisher information for meteorological variables introduction and feature selection in STLF was presented. By means of the approach, the main features can be extracted from the captured data and the cumulative effects of meteorological factors on STLF can be handled properly. The STLF model constructed by the methodology can always be updated with the latest data. When it was applied to real load data and the meteorological data, the predictive accuracy is still higher than traditional one even when the process features change rapidly.