A Fault Early Warning Method for Coal Mills Based on Causality and LSTM Model

Under the background of renewable energy constantly affecting the stability of the power grid, the peaking characteristics of thermal power generating units have become particularly important, which also puts forward high requirements for the safe and stable operation of various types of thermal power plant equipment. Efficient and convenient data-driven methods are gradually widely used in equipment operation and maintenance, and how to construct data-driven models with excellent performance is a matter of concern. In this paper, a fault early warning framework for coal mills based on causality and LSTM model is proposed, which consists of three steps: input feature selection, normal behavior modeling and fault early warning. Based on a real-world coal mills fault case, this paper discusses the effects of different feature selection methods, including expert experience, correlation and causality, on the modeling accuracy of LSTM models and the performance of early warning tasks. The results show that the LSTM model constructed based on causality has the highest prediction accuracy and better early warning performance, which provides a solution for practical engineering applications.