Research on Turbine Power Prediction Based on LSTM-KAN and Online Oil Monitoring Data

Large-scale machinery plays a significant role in industrial fields such as marine, aviation, and power production. However, due to the complexity of industrial production environments, equipment inevitably experiences wear and tear and may even malfunction. Power parameters are crucial indicators of equipment operating conditions. Using deep learning algorithms for trend prediction of these parameters can anticipate equipment health status and facilitate maintenance. Nonetheless, many traditional prediction methods perform poorly when handling complex nonlinear relationships and have limitations when dealing with high-dimensional data inputs and long sequence processing. In this paper, we propose a hybrid network called LSTM-KAN. LSTM, benefiting from its unique gating mechanism, excels at capturing long-term dependencies in time series. KAN, by improving the basic modules of MLP, effectively addresses issues like gradient vanishing, exploding, and low parameter efficiency in traditional neural networks. It possesses excellent capabilities for handling high-dimensional data and enhances model interpretability. This study establishes a predictive model using LSTM-KAN to link online oil monitoring data with equipment power and predict the equipment’s power output. Experimental comparison results show that the LSTM-KAN prediction model achieves an MSE of 2.6457 × 10−2 and an R2 of 8.3642 × 10−1, outperforming the LSTM prediction model, which has an MSE of 3.3225 × 10−2 and an R2 of 6.7933 × 10−1. This model enables long-sequence power prediction for equipment, enhancing prediction accuracy and reliability, and provides a new approach for the condition assessment of mechanical equipment.