Soft sensors have been widely employed in industrial processes to estimate process variables that are difficult to measure in real time. Taking into account process dynamics in soft sensor models is important not only for high predictability, but also for the interpretation of the models. One simple way of introducing dynamics is to use process variables X with time-delays. In this study, we propose a strategy of building soft sensor models for selecting appropriate variables X with time-delays, by incorporating ensemble learning into genetic algorithm-based process variables and dynamics selection (GAVDS). The construction of soft sensor models using simulation data sets and publicly available process operation data sets showed that the ensemble GAVDS method was superior to GAVDS in terms of model predictability. Furthermore, it appeared to select reasonable time delay regions for process variables X.