A Theory-Consistent Multi-Modal Fused Neural Network for Predicting Mixed Oil Length Range in Multi-Product Pipeline

Abstract In the operation of a multi-product pipeline, a precise and effective prediction model for the contamination amount range is crucial to guide the formulation of the cutting plan. Existing methods, however, fail to extract clear feature correlations and overlook the implicit physical knowledge of mixed oil, leading to impractical prediction results. Therefore, this study integrating latent feature representations and scientific theory into the neural network to design a theory-consistent multi-modal fused neural network for predicting the mixed oil length range. The proposed model establishes a comprehensive feature representation of multi-dimensional features of mixed oil to extract ample feature correlations through a multi-modality feature fusion module. Subsequently, a theory-guided customized neural network is designed to incorporate the scientific theory into the forward and backward propagation. Additionally, a physics-consistent loss function is established by incorporating range difference constraints and data mismatch, intending to prevent the model from producing physically unreasonable results. Validation on real-world cases suggests that the proposed model surpasses advanced techniques and recent achievements, reducing Root Mean Squared Relative Errors by 31% and 36% in lower and upper limit prediction, respectively. The sensitivity analysis of model components suggests that both the multi-modality feature fusion and physical principles are crucial for model improvement.