Multicycle and multimode are important features in fixed-bed reactors due to a manifold of reasons such as catalyst regeneration and equipment updates. Unfortunately, samples are not sufficient to establish an accurate model because of the frequent changes in the operating conditions. Moreover, a large amount of data from the historical cycle cannot be used directly due to different operating conditions. The online modeling of these processes faces significant challenges, such as lack of samples, nonlinearity, and multimode characteristics. To overcome this problem, an adaptive JIT-TL-SFA modeling approach is proposed by merging transfer learning (TL) and slow feature analysis (SFA) into just-in-time (JIT) learning. A novel time-space similarity measure criterion, which considers temporal relevance and spatial relevance to improve the performance of JIT, is proposed in this work. The strategy is implemented and tested on an acetylene hydrogenation process, and the results are presented and analyzed.