Résumé : Urban airflow and pollutant dispersion are strongly influenced by wind direction, which leads to distinct flow regimes characterized by different recirculation patterns, channelling effects, and plume trajectories. These regime changes are difficult to predict using global Reduced-Order Models that apply dimensionality reduction across all wind conditions. To address this limitation, this work develops an Expert-Based Reduced-Order Modelling framework that explicitly accounts for dependence on wind direction in urban environments. The proposed approach uses wind-rose maps to partition the wind-condition space into directional groups. Within each group, local Reduced-Order Models are constructed using Proper Orthogonal Decomposition for dimensionality reduction and Gaussian Process Regression to map wind conditions to reduced coefficients. A hard-gating strategy assigns new wind conditions to the appropriate expert during prediction. The framework is trained and validated using 200 Large Eddy Simulations of the MUST (Mock Urban Setting Test) configuration, spanning a wide range of inlet wind directions and friction velocities. Results show that wind-rose-based clustering significantly improves the compactness of the reduced bases and the predictive accuracy. Compared to the global Reduced-Order Model, the Expert-Based framework reduces prediction errors by more than 50% for velocity components and up to 75% for pollutant concentration. The proposed method enables fast and reliable predictions and is well suited for urban wind-engineering applications.