With the increasing popularity of electric buses (EBs), accurate estimation of the trip-level energy consumption of EBs has become increasingly essential. In this paper, an effective energy consumption estimation approach was proposed based on real-world operation data of EBs. Operation data of EBs from 3 different bus routes were collected and pre-processed to extract energy consumption-related features from various aspects such as traffic condition, environment, vehicle status and driving behaviour. The analyses of feature distribution, feature interaction and feature importance were then carried out. And the contributions of features to energy consumption were thoroughly analysed by Shapley value. Finally, different machine learning models were built and compared. The optimal results achieved an MAPE of 4.404% under 10-fold cross-validation, with an improvement of at least 29% over existing studies on the subject.