In the U.S., educational facilities consume a large amount of energy. Model predictive control schemes can improve the energy efficiency of educational facilities. Accurate and fast prediction of the cooling load is essential to performances of model predictive control schemes. Although many methods for the cooling load prediction were proposed, they are not suitable for educational facilities due to the lack of an efficient way to reflect the impact of internal activities on the cooling load. After analyzing the characteristics of cooling load of educational facilities, we proposed to use the day type instead of the day of the week as the input for the prediction. Then we constructed a Bayesian Network model based on that. To evaluate how the proposed inputs enhance the cooling load prediction, we also implemented the other Bayesian Network model with inputs recommended by the literature. To assess performances of those models, we performed a case study in which on-site measured cooling load and meteorological data was used for the training and testing. The results show that the Bayesian Network models can capture the trend of cooling load even with a limited size of training data. Replacing the day of the week by the day type can significantly improve the accuracy of cooling load prediction for educational facilities.