Physics-Guided Hybrid Intelligent Method for Predicting Mixed Oil Length in Multiproduct Pipelines

Abstract To ensure the reasonableness of the contamination oil cutting plan as well as to improve product quality and operation efficiency, accurately predicting the length of mixed oil in pipelines is crucial. However, due to the high-dimensional nonlinear characteristics in mixed oil multidimensional features and complex mass transfer mechanisms, existing research fails to achieve the desired accuracy and reliability resulting from the insufficient nonlinear fitting capabilities and neglect of mixed oil mechanisms. This work proposes a physics-guided knowledge-enhanced hybrid intelligent method (PG-KE-HIM) for accurate mixed oil length prediction. First, the input features are reconstructed based on mixed oil mechanisms and the optimal feature subsets are selected by comparing the predicting errors. Second, a deep neural network model is established to capture the intricate nonlinear relationships between the input attributes and the desired output. Subsequently, the empirical model is calibrated by historical data and correction coefficients to further enhance its predictive accuracy. Ultimately, the deep learning model is seamlessly integrated with the existing mechanism model through adaptive weights, enabling more accurate predictions that combine the strengths of both approaches. Results show that the proposed model offers more accurate predictions than existing methods, with RMSE, MAE, and MAPE being 101.603 m, 78.055 m, and 11.28%. To further elaborate, feature engineering based on domain knowledge and adaptive integrating deep learning models with mechanistic models are of great significance to enhance the interpretability and accuracy of prediction results.