Abstract: The heat transfer coefficient (HTC) is a key parameter to determine control precision in a cooling-controlled model. However， the correction for the HTC isn’t always stable and robustness by a traditional self-learning method. In order to solve such problems，the deep neural networks for the self-learning HTC were built based on the deep learning technology. The optimization of hyper parameter and the selection of algorithm in the network frame are studied. Therefore， the stability of the cooling control model can be greatly enhanced. The application of the 3500mm plate mill plant proves the better accuracy and robustness of the model， which can meet the requirements of actual on-site production.

Key words: deep learning, plate, ultra fast cooling, heat transfer coefficient, hit rate