Urban environments significantly affect atmospheric dynamics, influencing local climate, air flow patterns, and pollutant dispersion. Accurate modelling of urban flow physics is essential to mitigate adverse environmental effects and support sustainable urban design. However, high-fidelity computational fluid dynamics (CFD) simulations remain computationally expensive, while experimental measurements often provide limited spatial and temporal resolution. To address these limitations, modal decomposition techniques have emerged as powerful tools for extracting dominant flow structures and enabling reduced-order modelling. In this work, an initial database of 50 validated CFD simulations of simplified urban configurations under varying inflow and buoyancy conditions is analyzed using a Higher-Order Singular Value Decomposition (HOSVD) framework. Unlike classical SVD, HOSVD operates on high-dimensional tensor data. The proposed framework enables accurate reconstruction and interpolation of unseen flow conditions using a reduced set of modes through the application of machine learning and deep learning techniques. Results demonstrate that the methodology captures the dominant airflow structures, buoyancy-driven effects, and pollutant dispersion mechanisms with high fidelity while significantly reducing the dimensionality of the problem. Furthermore, the methodology is extended to a real urban neighborhood in Madrid to assess its predictive capability in realistic conditions. The predicted flow fields show strong agreement with CFD reference solutions and can be used as physics-informed initial conditions, reducing CFD convergence time and overall computational cost. These results demonstrate that HOSVD-based reduced-order modelling provides an efficient and robust approach for prediction and acceleration of CFD simulations in complex urban environments, with potential applications in urban climate modelling, pollution assessment, and real-time environmental analysis.