Building Performance Simulation (BPS) is a potent and commonly utilized method to assess and improve building design, operation, and overall energy performance. Commonly used physics-based BPS tools often entail a high computational cost and require a significant amount of technical data and multiple assumptions for energy predictions. While data-driven approaches are contingent upon the availability of large enough volumes of high-quality data for training the prediction models. Surrogate models, which combine physics-based simulations with components of statistical or machine learning techniques have demonstrated their ability to harness the advantages of both physics-based and data-driven models to achieve superior prediction results. In this study, a surrogate model is utilized to predict the cooling energy consumption for a mixed-use district. A hybrid approach is employed, which combines machine learning techniques with representative building physics-based archetypes. High-rise building archetypes, representative of the building stock in the district, are modeled using EnergyPlus software. A parametric simulation is then automated using RStudio by modifying crucial parameters related to the building envelope characteristics, building geometry, and building operational parameters, including occupant behavior. The proposed methodology offers a solution for a more accurate representation of occupant behavior in machine learning surrogate building energy models. The artificial dataset generated is then used to train three neural network models, one for each of the predominant building typologies available in the district. The predictions made by the machine learning models are then benchmarked against unseen EnergyPlus results, and the prediction performance of the models is validated using findings in the literature.