Explainable machine learning methods for predicting electricity consumption in a long distance crude oil pipeline

Area of Science:

• Energy Systems Engineering
• Machine Learning Applications
• Pipeline Transportation

Background:

• Accurate electricity consumption prediction is crucial for optimizing energy use and costs in crude oil pipelines.
• Traditional machine learning models struggle with nonlinear features, accuracy, and interpretability in this domain.

Purpose of the Study:

• To develop a novel, accurate, and interpretable electricity consumption prediction model for crude oil pipeline transportation.
• To address limitations of traditional machine learning algorithms in considering influencing factors and extracting nonlinear features.

Main Methods:

• Integration of Grid Search (GS) for hyperparameter optimization and Extreme Gradient Boosting (XGBoost) for predictive modeling.
• Application of SHapley Additive exPlanations (SHAP) to enhance model interpretability by quantifying parameter contributions.

Main Results:

• The proposed GS-XGBoost model significantly outperformed other benchmark models (MLP, SVM, ELM, RF, GBRT).
• Achieved a mean absolute percentage error (MAPE) of 4.1% and a coefficient of determination (R²) of 0.98.
• SHAP analysis identified key predictors: daily transport volume, pump-out pressure, and station pressures/temperatures.

Conclusions:

• The GS-XGBoost model offers superior predictive performance and interpretability for pipeline electricity consumption.
• The findings support improved energy efficiency, cost optimization, and digital transformation in pipeline operations.