A Hybrid Modeling Method for Efficiency Prediction of Centrifugal Pumps at Crude Oil Pipeline Stations

Abstract

Centrifugal pumps are critical energy-consuming assets in crude oil pipeline stations. The precise prediction of their operational efficiency is vital for energy conservation, operational stability, and predictive maintenance. However, traditional modeling approaches face significant limitations. Standalone mechanistic models are based on ideal design specifications. These models fail to capture performance degradation that results from physical wear and time-varying operational conditions. Conversely, purely data-driven models, such as the standard Artificial Neural Network (ANN), often exhibit poor generalization and unreliability. This occurs when they are confronted with data-scarce operating regimes not adequately represented in the training data. To address these challenges, this paper proposes a novel hybrid model architecture: the Physics-Data Collaborative Fusion Neural Network (PDC-FNN). This model features a parallel dual-channel structure. A data-driven channel processes raw sensor data to extract complex operational patterns. In parallel, a physics-guided channel interprets prior knowledge from an ideal mechanistic model to provide a theoretical baseline. We conduct a comprehensive performance comparison between the PDC-FNN, a traditional mechanistic model, and a standard ANN. The comparison uses a real-world industrial dataset characterized by imbalanced operating conditions. The experimental validation is performed under both data-rich (basic prediction) and data-scarce (generalization) scenarios. The results demonstrate the clear superiority of the proposed method. The PDC-FNN maintains high prediction accuracy across all tested conditions. In contrast, the standard ANN’s performance collapses in the data-scarce regime. These results confirm that the PDC-FNN’s unique structured fusion approach provides an effective and robust solution. It enables the development of high-precision, reliable predictive models for complex industrial systems.