From Complexity to Clarity: Structural Process Knowledge-Informed Neural Network for Alumina Concentration Distribution Prediction

Maintaining an optimal alumina concentration distribution is crucial for ensuring stable operation and reducing energy consumption in aluminum electrolysis cells. Due to the strong coupling of multiple physical fields within the electrolysis cell, the rapid and accurate prediction of alumina concentration distributions has long been a formidable challenge. Furthermore, the harsh industrial environment only allows for the acquisition of alumina concentration data at limited locations, hindering the construction of precise predictive models for the alumina concentration distribution. Moreover, the presence of measurable disturbances, such as feeding, can exert additional influences on the alumina concentration, necessitating traditional methods to collect new data and reconstruct models. To address these issues, this article proposes a structural process knowledge-informed neural network for alumina concentration distribution prediction with desirable precision and high solving efficiency. Specifically, considering the alumina concentration distribution is primarily driven by electrolyte convection and secondarily by electrolysis consumption, pretraining modules for flow and concentration field are designed to explicitly modify the network structure, thereby strengthening the relationship between inputs and process state variables, which significantly enhance prediction efficiency and accuracy. Subsequently, a loss function incorporating process governing equations is designed, which effectively constrains the relationship between inputs and process state variables to adhere to physical principles. This enables accurate concentration prediction at any location in the field, even with a limited number of spatial training data. Finally, a training method for alumina concentration prediction under varying initial concentrations is introduced, incorporating the initial concentration and a fixed time step into the network inputs, thereby eliminating the need for traditional methods to reconstruct models for continuous predictions under different initial conditions. Extensive experiments demonstrate that our method achieves significant performance improvements compared to some state-of-the-art methods.