To address the issue that vehicle behavior prediction based on in-vehicle monitoring cameras may infringe on the privacy of drivers and other road participants, this study develops a bus activity prediction model with vehicle motion and behavioral operation data as input. First, a naturalistic urban bus driving experiment was conducted, collecting vehicle motion and driver operation data via the CAN bus. Subsequently, segments corresponding to station entry, station exit, intersections, turning, and lane changing were selected. The phase space reconstruction algorithm was used to map the time-series data into a high-dimensional space, thereby generating an RGB image dataset. Finally, an E-bus Vehicle Behavior Prediction Model (E-VBPM) was established based on the ConvNeXt network. The results indicate that the developed E-VBPM achieved an accuracy of 84.62% in predicting driving activities, representing an improvement of approximately 6.8% over machine learning models that utilize time-series data inputs. These findings support the development of more intelligent on-board systems for electric buses, enabling better identification of vehicle operating modes and enhanced driver assistance.