This paper considers an auction-based parking reservation problem where a parking management platform is the auctioneer and the drivers are bidders. The platform is in charge of multiple homogeneous parking spaces. A winner may leave earlier or occupy the parking space longer than the time he has reserved. The phenomena are known as (ex post) demand disturbances, which can occur only after the last auction terminates. The platform may penalize or compensate a driver who causes demand disturbance. Besides, investigation is conducted into three types of driver behaviors, namely, gain/loss neutrality, loss aversion, and gain seeking, and the reference effects are examined. An effective multi-stage Vickrey–Clarke–Groves (MS-VCG) auction mechanism is raised. Expect for the disturbance makers, the MS-VCG auction is capable to achieve allocative efficiency, incentive compatibility, and individual rationality. As shown by computational results, in the absence of reference effect, both penalty and compensation rise with the number of bidders. If drivers are gain/loss neutral, the average utility of winners reaches the highest, while the penalty and compensation reach the lowest, as compared to loss-averse and gain-seeking drivers. The total VCG payment received by the auctioneer increases with the reference price. Finally, based on our proposed model, the platform has the ability to deal with the demand disturbances by holding a certain number of parking spaces that are not auctioned for reservation.