Ozone prediction, a key role for ozone pollution control, is facing the following challenges, i.e., the complex evolution trend of ozone, the cross-interference phenomena between ozone and other pollutants, and the low-quality monitoring data. To overcome the above challenges, we propose a multi-source and multivariate ozone prediction model based on fuzzy cognitive maps (FCMs) and evidential reasoning theory from the perspective of spatio-temporal fusion, termed as ERC-FCM. In this framework, an FCM-based prediction model is introduced to solve the ozone forecasting problem. Inspired by the multivariate time series forecasting, a multivariate ozone prediction problem is modeled as an FCM learned by the real-coded genetic algorithm, in which each node denotes a variable (pollutant). Thus, both the complex evolution trend of ozone and the cross-interference phenomena can be reflected by the FCM. Further, we propose an ensemble theoretical framework based on evidence reasoning theory and the matrix 2 norm. This theoretical framework relieves the negative factors from the low-quality monitoring data and improves the prediction accuracy when facing multi-source and multivariate time series. The performance of ERC-FCM is validated on two real-world datasets. The experimental results demonstrate that our method yields the best prediction performance by comparison with the other classical FCM-based methods on mean absolute error (MAE), mean square error (MSE), and root mean square error (RMSE). In addition, the Friedman test and Nemenyi test show that ERC-FCM gets relatively better prediction accuracy than other models.