Organic-inorganic hybrid perovskites have promising applications in solar cells and other optoelectronic devices due to their superior physical and chemical properties. A band gap is a key physical characterization that is related to the efficiency of solar energy conversion. We performed machine learning for the band gap of hybrid perovskites, and investigated the influence of structural features based on the Voronoi method on the model accuracy. The results show that compared to machine learning with only element features as an input, the accuracy of the band-gap models can be improved when the Voronoi structural feature is included in all the three methods of symbolic regression (VS-SISSO), artificial neural network (ANN) and random forest (RF). In particular, the Voronoi structural feature is of vital importance in the VS-SISSO model for its nature of being simple explicit expressions. The models from the three machine learning methods have comparable prediction accuracies, and the VS-SISSO model has better transparency and interpretability. According to the feature importance analysis, the Voronoi structural feature is the most important among all the input features.