Riassunto analitico
In the pursuit of improved energy efficiency in industrial plants, accurate prediction of pump performance plays a pivotal role. Centrifugal pumps, essential to a myriad of industrial processes, are traditionally modeled using affinity laws, which provide mathematical relationships between pump speed and performance parameters such as flow rate, head, and power. However, these laws rely on manufacturer-specified data and are often limited when predictions are required at rotational speeds significantly different from those provided.
This thesis investigates the enhancement of pump performance predictions by integrating data-driven correction factors with traditional affinity laws. A hybrid modeling approach was developed that combines classical engineering methods with advanced machine learning techniques. The study explores several predictive models, including baseline Dense Feedforward Neural Networks (DFNN), refined DFNNs optimized through extensive hyperparameter tuning, Physics-Informed Neural Networks (PINN), and ensemble methods such as XGBoost, Random Forest, bagging, boosting, and stacking. Each model was trained on a dataset derived from manufacturer specifications and performance graphs, which included features such as flow rate, head, power, and design specifications. After preprocessing, the models were evaluated on the original scale of the target variables using performance metrics such as Mean Squared Error (MSE), Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), Mean Absolute Percentage Error (MAPE), and a custom improvement metric.
Results demonstrate that incorporating data-driven correction factors significantly improves prediction accuracy over traditional affinity law estimates. Notably, ensemble methods, particularly XGBoost and Random Forest, achieved the lowest absolute errors and the highest custom improvement metrics, indicating their strong ability to capture the nonlinear relationships inherent in pump performance data. Additionally, regression and residual analyses confirm that the correction factors substantially reduce systematic discrepancies.
This research lays a robust foundation for further advancements in predictive modeling for industrial pump systems and offers promising directions for future work, including adaptive correction strategies, enhanced feature engineering, and real-time integration with industrial monitoring systems.
|
