Аннотация:This study presents preliminary results of our research, aimed to development of a technique for downscaling the meteorological parameters in urban areas based on machine learning (ML) methods exemplified by application to the Moscow megacity. Firstly, we consider a problem of approximating the temporal variations of the urban heat island (UHI) intensity observed at a specific urban location based on predictors characterizing large-scale weather patterns. Such predictors are derived from rural weather observations and ERA5 reanalysis data. Several ML-based models are compared, including Random Forests, Gradient Boosting, Support Vectors and Multi-layer Perceptron. These models, trained on a 20-year dataset, successfully capture the diurnal, synoptic-scale and seasonal variations of the UHI intensity based on predictors derived from either observations of reanalysis. Evaluation scores are further improved when using both types of predictors simultaneously and involving additional features characterizing[1] temporal dynamics of predictors (change rates and moving means during 3-12 hours prior to target case). The best scores are achieved with boosting models, firstly CatBoost Regression. Surprisingly, the best ML-based models forced only by reanalysis data perform better than comprehensive mesoscale model COSMO, supplied by urban canopy scheme, detailed city-descriptive data and used for dynamical downscaling of the same reanalysis. Secondly, we present an architecture and initial tests of the ML-based technique for simultaneous approximation of temporal and spatial variability of the UHI intensity. Here, we use reanalysis-based predictors characterizing large-scale weather patterns (time series), and a set of additional predictors characterizing land cover and city-descriptive parameters (time-invariant 2D fields). To train ML-based models, we use the in-situ observations as well as output fields from high-resolution simulations with COSMO model with urban canopy scheme, forced by same reanalysis and landcover data as used as predictors. The study is supported by Non-commercial Foundation for the Advancement of Science and Education INTELLECT.