Vehicle-to-grid (V2G) technology utilizes scheduling models to generate decision schedules that enable electric vehicles (EVs) to participate in grid management in an orderly manner, which can achieve efficient peak shaving and valley filling. The federated learning approach can be used to train the scheduling model under the condition that the charging stations are unwilling to share the raw data. Therefore, selecting training labels that align with the interests of multiple parties and ensuring the correctness of model parameter aggregation are crucial for V2G scheduling decisions. In this paper, a trusted federated learning method for V2G scheduling is proposed. First, a trusted federated learning architecture for real-time V2G scheduling is constructed, which includes three components: a label generation module, a verifiable federated learning module, and a real-time scheduling module. Next, a real-time scheduling label data generation model is proposed, considering the interests of EV users, operators, and grid-side load fluctuations, and a dynamic updating method for scheduling model labels is designed. Furthermore, a secure record and verification method for model parameter aggregation is proposed to ensure the correctness of federated learning model parameter aggregation. Finally, the time overhead, storage overhead, and Gas cost of the label data generated by EVs in three types of charging time periods and the proposed verification method are analyzed. Numerical results indicate that the proposed label model exhibits optimal values for EV users, operators, and grid-side load fluctuations, with the time overhead for constructing the aggregation tree reaching the millisecond level. Compared to traditional verification methods, the Gas cost of aggregating and verifying using the smart contract is significantly reduced. Thus, the proposed trusted federated learning method aligns with the interests of multiple stakeholders in the power grid and demonstrates superior performance.