With the development of autonomous vehicular technologies, the execution tasks become more memory-consuming and computation-intensive. Simultaneously, a certain portion of tasks are latency-sensitive, such as collaborative perception, path planning, collaborative simultaneous localization and mapping, real-time pedestrian detection, etc. Because of the limited computation resources inside vehicles and restricted transmission bandwidth, edge computing can be an effective way to assist with the tasks execution. Considering from the perspective of business, the reservation or subscription cost is cheaper than real time requests. In order to minimize the expense of consuming edge services, the desirable situation is to reserve the resources as much as needed. However, the configuration of vehicular network is variational in practice due to the diversity of road maps, different time range like peak time and off-peak time, and the various task types, which makes it challenging to figure out a general machine learning model that is suitable for any case. Therefore, to predict the resource consumption in edge nodes accurately in different scenarios, we propose a two-stage meta-learning based approach to adaptively choose the appropriate machine learning algorithms based on the meta-features extracted on database. Besides, due to the deficiency of dataset for edge resource consumption, we program in game engine unity to generate the 3D model of Manhattan area. Meanwhile, we change the factors like different road maps and number of vehicles so as to get closer to practices. In the evaluation part, we adopt root mean square error, mean absolute percentage, and mean GEH as evaluation metrics to assess the performance of each model. Also, a quantitative analysis for the total cost and waste is also conducted. Eventually, we can find that the proposed meta-learning based method outperforms the non-meta ones.