Deep reinforcement learning (DRL) offers a sophisticated approach to developing intelligent traffic light controllers, with various DRL methods having been proposed in recent years. However, much of the research is limited to laboratory, simulation settings, and simple network structures, leaving their effectiveness in real-world scenarios largely unproven. In this work, we construct a large traffic network based on real-world conditions, in Birmingham, UK, and assess its performance in a wide variety of simulated realistic traffic scenarios. We assess performance improvements both locally (on a per junction basis) and globally (across the network) when applying DRL-trained controllers to key junctions. Our findings highlight the substantially positive impact our DRL-trained controller has under imbalanced traffic conditions (e.g., high traffic volume on certain routes and not on others, during peak hours), where it outperforms traditional pressure-based controllers in improving local traffic performance. However, they do not consistently enhance global performance and indicate the need for more research to achieve this aim.