High-efficiency pump model dataset: AI-optimization benchmark and CFD-experimental validation

Abstract

Pumps play a crucial role in fluid transportation, accounting for nearly 10% of global electricity consumption, and related energy-saving research is a key global concern. In practical applications, the traditional optimization based on genetic algorithms requires a huge search space and yields unsatisfactory results. By contrast, AI algorithms centered on neural networks serve as the current industry mainstream based on the training-feedback-prediction model. Evidently, a rich, accurate training set is critical. This paper, for the first time, presents 123 engineering-validated, publicly parameterized high-efficiency centrifugal pump models retrieved from 121 references (20+ English, ∼100 Chinese publications) and constructs a rigorous and reliable high-efficiency model dataset spanning a specific speed range of 35–500. Further, a parametric predictive framework developed by Python, where the back-propagation neural networks (BPNNs) and Long Short-Term Memory (LSTM) are employed for training and predictive modeling. Benchmarks show that BPNN has better adaptability and accuracy. On this basis, case studies on pumping machinery were conducted, and an optimization approach validated via CFD-experimental coupling yielded a statistically significant performance leap (Δη/η > 9.23%, ΔH/H <3.02%) by just 23 samples. This method greatly reduces the number of samples as well as computation cost as compared to traditional genetic algorithm, which shows great potential of energy saving for fluid transportation process.