Resolving the wind profile in an urban canyon environment means dealing with the turbulent nature of the stream and the presence of non-negligible flux exchanges with the atmosphere inside the canopy, making any deterministic model solution computationally very intensive. In this paper, a learning-from-data method is explored, which is able to predict the wind speed in an urban canyon at different heights, given a minimal set of input features. The experimental location is provided by a street canyon located at the Swiss Federal Institute of Technology campus in Lausanne, equipped with several measuring stations to record data at high temporal resolution. Different machine learning approaches are compared in order to predict the wind speed in two directions and at different heights inside the urban canyon: an optimized Ridge Regression outperforms the Random Forest algorithm. We find particularly high accuracy in predicting the wind speed in the highest part of the canyon. None of the proposed algorithms however is able to model in an accurate way the variation of the wind speed close to the ground.