We consider a hierarchical network that consists of mobile users, a two-tiered cellular network (namely small cells and macro cells) and central routers, each of which follows a Poisson point process (PPP). In this scenario, small cells with limited-capacity backhaul are able to cache content under a given set of randomized caching policies and storage constraints. Moreover, we consider three different content popularity models, namely fixed content popularity, distance-dependent and load-dependent, in order to model the spatio-temporal behavior of users' content request patterns. We derive expressions for the average delay of users assuming perfect knowledge of content popularity distributions and randomized caching policies. Although the trend of the average delay for all three content popularity models is essentially identical, our results show that the overall performance of cached-enabled heterogeneous networks can be substantially improved, especially under the load-dependent content popularity model.