Multi-objective optimization of hot water circulation pump using machine learning model and non-dominated sorting genetic algorithm

To enhance both the hydraulic efficiency and noise performance of a hot water circulation pump, key design variables were identified using the Plackett–Burman experimental design. These variables include the impeller outlet width, blade wrapping angle, and volute base circle diameter, which significantly influence the pump’s performance. A database of 60 optimization samples was generated through the Latin hypercube sampling method, enabling the development of predictive models using eXtreme Gradient Boosting (XGBoost), back propagation neural network, and least absolute shrinkage and selection operator algorithms. Among them, XGBoost demonstrated superior accuracy in predicting the pump’s performance. For optimization, non-dominated sorting genetic algorithm III was employed, yielding an optimal configuration with parameters D3 = 429 mm, φ = 119°, and b2 = 36.9 mm. Compared to the original model, this optimized pump performs better hydraulically over the flow range from 0.8Q to 1.2Q, which achieves a reduction in entropy production by 5.32% in the impeller and 3.33% in the volute. Noise levels also improved substantially after optimization. The high sound power region near the blade working surface disappeared, and the total sound pressure level decreased by 3.02 dB.