2GJ of energy is required to bake one ton of carbon anodes which are heat-treated in the anode baking furnaces. Computational fluid dynamic (CFD) modeling is useful in conducting coupled transient heat transfer, turbulent fluid flow, and combustion simulations. However, due to the huge temporal and spatial domains, it is difficult and expensive computationally to use CFD models for the evaluation of the overall furnace operation. As a part of quality control in most of the modern aluminum smelters, large flue-gas and anode temperature measurements are available. The present study applies an artificial neural network-based method to better exploit these large datasets and gain new physical insight in a cost-effective manner. A shallow neural network is considered consisting of an input layer, a hidden layer, and an output layer. The data is divided into a training set (70%) and a validation set (30%), to avoid overfitting of the data. It is remarked that there is a good agreement between the fitted data and targeted values for both the validation set as well as the training set. For the case study, 147 epochs are required to reach the best validation performance, which is 84.9. The error between predicted and target values are mostly between 20C, which indicates a high accuracy in the prediction level. The neural network-based model can be effectively integrated into the anode baking process models to estimate anode baking uniformity more precisely. The methodology presented in the current research can be extended to simulate the entire anode baking process, preheating, firing and cooling sections, cost-effectively.
Keywords Anode baking furnace, artificial neural network, anode baking, flue-wall