Air-conditioners (AC) usually consume the most electricity among all of the auxiliary components in an electric bus, over 30% of the battery power at maximum. On-board passengers carried by the electric bus are important but random heat sources, which are obsessional disturbances for the cabin temperature control and energy management of the AC system. This paper aims to improve the AC energy efficiency via passenger amount variation analysis and forecast in a model predictive control (MPC) framework. Three forecasting approaches are proposed to realize the passenger amount variation prediction in real-time, namely, stochastic prediction based on Monte Carlo, radial basis function neural network (RBF-NN) prediction, and Markov-chain prediction. A sample passenger number database along a typical bus line in Beijing is built for passenger variation pattern analysis and forecast. A comparative study of the above three prediction approaches with different prediction lengths (bus stops in this case) is conducted, from both the energy consumption and temperature control perspectives. A predictive AC controller is developed, and evaluated by comparing with Dynamic Programming (DP) and a commonly used rule-based control strategy. Simulation results show that all the three forecasting methods integrated within the MPC framework are able to achieve more stable temperature performance. The energy consumptions of MPC with Markov-chain prediction, RBF-NN forecast and Monte Carlo prediction are 6.01%, 5.88% and 5.81% lower than rule-based control, respectively, on the Beijing bus route studied in this paper.