The paper considers the possibility of thunderstorm forecasting using only dynamical and microphysical parameters of the cloud, simulated by the 1.5D model with further processing by machine learning methods. The problem of feature selection is discussed in two aspects: selection of the optimal values of time and height when and where the output model data are fixed and selection of fixed set of the most representative cloud parameters (features) among all output cloud characteristics. Five machine learning methods are considered: Support Vector Machine (SVM), Logistic Regression, Ridge Regression, boosted k-nearest neighbour algorithm and neural networks. It is shown that forecast accuracy of all five methods reaches values exceeding 90%.