Benchmarking Linear vs. Non-Linear Predictive Models for Energy Demand Forecasting in Electrical Motor-Driven Compressors

Abstrak

In the LNG regasification plants, among the most energy-demanding components are the electrical motor driven compressors. Artificial Neural Networks (ANNs) and Extreme Gradient Boosting (XGBoost) feature superior flexibility in modelling nonlinear and dynamic compressor behaviour, unlike the conventional methods such as Multiple Linear Regression (MLR), which is limited by its assumption of linearity. This study focuses on data-driven techniques for compressor power consumption prediction, in which MLR, ANN and XGBOOST are compared. Here, real operational data collected from two Boil-off Gas (BOG) and two Regasification Terminal Export Compressors (RGTEC) were considered to develop and evaluate the models. The findings suggest that non-linear machine learning models provide superior predictive performance in comparison to those of the traditional linear approach. The highest prediction accuracy for most compressors is achieved using the ANN model, with a n R2 values of 94.6%, 99.64% and 99.64% for BOG-A, BOG-C and RGTEC-A, respectively. Meanwhile, the best performance was found for XGBOOST for RGTEC-B, with a R2 value of 94.17% and a significantly lower RMSE value. Contrary to that, MLR produced lower accuracy in several cases, particularly when subjected to complex operating conditions, in which linear assumptions were inadequate to capture system dynamics. In addition, this research features one of the first direct comparisons between linear and nonlinear models applied to real-time compressor data, unlike past research that depends on simulations or single-method analysis. It emphasizes the practical advantages of ANN and XGBoost for data-driven energy forecasting and operational optimizations in the gas industry.

Kata kunci

• Load Forecasting
• Deep Learning
• Multiple Linear - Regression
• Electrical Compressors