Abstract: Abstract: Hydraulic pressure drop prediction is very important for production regulation of oil pipelines， and current machine learning methods regard pressure drop prediction as a regression problem， however， pipeline hydraulic calculation is affected by many factors， and the fixed weights obtained from the training set by traditional machine learning methods are difficult to generalize to more test samples or real engineering scenarios. This paper proposes a hydraulic pressure drop regression prediction method， Regformer， which introduces a sparse attention mechanism into the regression task， designs a smoothing probability method based on multi-headed attention， and incorporates a feature projection mechanism. In a comparative experimental analysis with seven mainstream methods on 10 public data sets， qualitative experiments show that Regformer has good fitting ability for local mutations; experiments on hydraulic pressure drop prediction show that the self-attentive method has significant advantages for regression tasks with multivariate uncertainty， especially for extreme cases reflecting the importance of adaptive regression parameters， and Regformer achieves better performance than Transformer with less computation， verifying the superiority of the proposed sparse attention and adaptive feature projection for the hydraulic pressure drop prediction task.

Key words: Key words: hydrodynamic prediction, Transformer, Regformer, Self-attentive