When cooling and exhaust air flows in air-cooled datacenters mix, the energetic efficiency of the cooling operations drops. One way to prevent this mixing of happening is by augmenting the air tightness of the hot and cold aisles; this, however, requires installing opportune hardware that may be expensive and require time consuming installations. Alternatively, one may try to minimize cooling and exhaust air flows mixing by opportunely controlling the speeds of the fans of the Computer Room Air Handling (CRAH) units so that the distribution of the air pressure field within the computer room is favorable. Implementing this type of flow control requires both detecting when there actually is some type of flow mixing somewhere, plus understanding how to operate the cooling infrastructure so that these mixings do not happen. To this aim, there is the need for models that can both help deciding whether these mixing events occur, plus designing automatic control strategies for reducing the risks that they will happen. In this manuscript, we propose an ad-hoc methodology for the data-driven derivation of control-oriented models that serve the purposes above. The methodology is built on classical Prediction Error Method (PEM) approaches to the system identification problem, and on laddering on the peculiarities of the physics of the phenomena under consideration. Moreover, we test and assess the methodology on a industrial-scale air-cooled datacenter with an installed capacity of 240 kW, and verify that the obtained models are able to capture the dynamics of the system in all its potential regimes.