基于径向基神经网络模型预测CuO-ZnO/(乙二醇-水)混合纳米流体导热系数

采用两步法制备CuO-ZnO质量分数为0~3 wt%的CuO-ZnO/(乙二醇(EG)-水)混合纳米流体。其中，纳米颗粒CuO和ZnO质量比为50∶50不变，基液混合比(EG与去离子水的质量比)变化范围为20∶80~80∶20，分析其导热系数随温度(25~60℃)及基液比的变化规律。然后，以质量分数、温度及基液混合比为自变量，导热系数为因变量，采用径向基神经网络(RBFNN)模型预测导热系数，并与反向传播神经网络(BPNN)模型和多元线性回归(MLR)模型的预测值对比。结果表明，CuO-ZnO/(EG-水)纳米流体导热系数随温度的升高呈非线性增大，当CuO-ZnO质量分数为3 wt%及基液混合比为20∶80时，其导热系数与纯基液相比增大了14.03%~23.47%；但随着基液中的EG含量增大，导热系数非线性下降。总之，CuO-ZnO/(EG-水)纳米流体的导热系数受粒子随机运动和温度变化呈非线性变化。采用RBFNN模型预测CuO-ZnO/(EG-水)纳米流体的导热系数，其结果与BPNN模型和MLR模型对比，RBFNN模型性能评价指标均方根误差(RMSE)、平均相对百分比误差(MRPE)及误差平方和(SSE)更接近于0，多元统计系数R2更接近于1，说明RBFNN模型预测导热系数的精度更高，能够较好地表征不同参数对导热系数的影响，为CuO-ZnO/(EG-水)纳米流体的热物理性能参数的预测提供了一种有效的数据驱动建模方法。 

Prediction of thermal conductivity of CuO-ZnO/(ethylene glycol-water) hybrid nanofluids based on radial basis neural network model

The CuO-ZnO/(ethylene glycol(EG)-water) nanofluid with CuO-ZnO mass fractions of 0–3 wt% were prepared by two-step method. First, the mixing mass ratio of nanoparticles CuO to ZnO was fixed at 50∶50, and the mass ratios of EG to deionized water varied from 20∶80 to 80∶20. Variations of thermal conductivity with temperature (25–60℃) and mixing mass ratios of base liquids were studied. Second, the radial basis function neural network (RBFNN) model was used to predict the thermal conductivity. In the model, the mass fraction, temperature and mixing mass ratios of base liquids were considered as independent variables while the thermal conductivity is the dependent variable. The predicted values were compared with the values predicted by back propagation neural network (BPNN) and multiple linear regression (MLR). The results show that the thermal conductivities of CuO-ZnO/(EG-water) nanofluid increase nonlinearly with the increase of temperature, while decrease with the increase of mixing mass ratios of base liquids. Compared with the thermal conductivity of the base fluid, it increases from 14.03% to 23.47% at the mass fraction of 3 wt% and the mixing mass ratio of base fluid of 20∶80. The thermal conductity of CuO-ZnO/(EG-water) changes nonlinear with random motion of nanoparticles and temperature. The results obtained by RBF model are more precise than that predited by BPNN model and MLR model. The model evaluated indexes of root mean square error (RMSE), mean relative percentage error (MRPE) and sum of squared error (SSE) are closer to 0 and duostatistical coefficient of multiple determination R2 is closer to 1 indicating that the RBFNN model can accurately predict the thermal conductivity. It can also be used to characterize the effect of various parameters on the thermal conductivity. The results offer an effective method to establish the data-driven model to accurately predict the thermal conductivity of CuO-ZnO/(EG-water) nanofluid.