The designed capacity of cooling supply is much higher than demand of medium in a circulating cooling water system, which leads to the great waste of electricity energy. In order to keep balance between cooling demand and supply, and improve the energy efficiency of the circulating cooling water system, a minimum differential pressure control scheme for the multi-process medium temperature target of the circulating cooling water system is presented, and deep learning is applied to the study of the medium temperature prediction. And a temperature prediction method for media based on improved stacked auto encoders(ISAE) is proposed. Firstly, the industrial data is cleaned. Then, multiple autoencoders are stacked, and the “greedy unsupervised pre-training layer by layer” method is used to train network parameters. The root mean square back propagation(RMSProp) optimization is deployed to fine-tune network parameters to reduce the possibility of falling into local optimums. Finally, the process medium temperature prediction model is obtained using off-line training of the historical operation data of a chemical plant. Compared with the results of shallow neural networks and unmodified SAE methods, the prediction accuracy of the proposed ISAE method is high, the mean absolute percentage error(MAPE) is only 0.85%, and the generalization ability is better than the unmodified SAE algorithm.