Modelling of Forecasting Kiln Tail Temperature in Lime Calcination Process using Modified Physics-Informed Networks

In this work, we proposed a modified physics-informed long-short-term memory (LSTM) network for modelling the lime rotary kiln dynamics, which can maintain the physical information about certain inputs and outputs by inte-grating the mechanism during the learning process of neural networks to achieve reasonable control-oriented prediction. In the proposed method, a one-dimensional mechanism is firstly constructed based on the direction of the rotary kiln roll axis, which includes thermal exchange, lime generation, and combustion models. Utilizing these mechanism models, the constraints of certain physical inputs-output relationships are imposed on corresponding paths of the network during the LSTM learning process, while additional data is produced and employed for extending the generalization of the learned LSTM. Both the mechanism model and LSTM are updated synchronously and respectively. Finally, the experimental data demonstrates the effectiveness of the mechanistic model and the accuracy of the modified physics-informed LSTM model.