The accurate assessment of the demand response potential of air conditioning loads is a crucial foundational task for effectively scheduling their participation in demand response. To address the issue of low prediction accuracy caused by the traditional deep learning methods’neglect of the time-sequential distribution differences in real-world air conditioning loads, the concept of transfer learning to the time dimension is extended. The phenomenon of time-sequential distribution drift in air conditioning load time series is analyzed by drawing an analogy to covariate shift in transfer learning. Based on this, two time-sequential migration strategies, time-sequential distribution matching and time-sequential similarity quantification, are proposed. These strategies are integrated into the traditional recurrent neural network（RNN）architecture to build an adaptive RNN air conditioning load prediction model, thereby improving the prediction accuracy in real-world scenarios. Finally, an air conditioning load demand response potential evaluation method is proposed based on the overall approach of predicting the load values before and after the response and the adaptive RNN air conditioning load prediction model. Comparative experimental analysis on real datasets shows that this method can significantly improve the prediction accuracy of demand response potential over existing methods, thus providing effective reference for the demand response scheduling decisions of the grid dispatch center.