Federated learning is a distributed learning paradigm that enables learning on multiple clients to solve data privacy problem. In intelligent transportation system, fuel consumption and emissions are affected by many dynamic parameters, such as seasons, weather, driving behavior, therefore, it is necessary to predict them. However, fuel consumption and emissions reflect privacy characteristics, hence, ensuring privacy while ensuring the prediction effect has become a challenge. In order to solve this problem, this study propose a multi-variable input traffic model based on federated learning, which predict fuel consumption and emissions based on speed, acceleration, temperature and seasons and apply real-world vehicle history data for simulation. To avoid affecting predicting effect, autoencoder is introduced to traffic model to remove outliers. The results show that the MSE after the outlier removal is at least 10% lower than the raw data, and this traffic model compared with the existing non-federal learning methods, the prediction loss value in MAE in winter only increase by about 0.2%-3.8%, the MSE in summer decrease by about 3%-17.8%. It is proved this model still have a favorable prediction performance while protecting privacy.

Federated learning;fuel consumption;emissions;autoencoder;multi-variable