Understanding airflow within urban areas is crucial for ensuring pedestrian comfort, optimizing building ventilation, managing air quality, and evaluating how wind affects structures. Computational Fluid Dynamics (CFD) is a recognized method for simulating wind flow in cities. However, its computational demands related to high-fidelity schemes, such as large eddy simulations (LES), pose a significant challenge, particularly when dealing with probabilistic analysis, risk assessment, and real-time predictions. Moreover, despite the availability of faster CFD simulations like Reynolds-averaged Navier–Stokes (RANS), these low-fidelity (LF) predictions are often inaccurate for simulating wind flow in urban areas. To address these challenges, this study proposes a novel hybrid machine learning model comprising a dimensionality reduction technique and a long short-term memory (LSTM) network. Specifically, proper orthogonal decomposition (POD) is used for dimensionality reduction, and then the time-dependent POD coefficients are obtained via direct projection of the LF and HF signals onto the identified POD basis. An LSTM network is subsequently trained to map the LF POD coefficients to their corresponding HF POD coefficients. To demonstrate the performance of the proposed approach, a simplified case study involving wind flow in an urban area is presented. The analysis confirms that the introduced framework can rapidly and accurately predict HF urban flow.