The work aims to use different artificial neural network algorithms to predict the holding time of cold chain incubator under different cold storage agent parameters in order to find the most suitable artificial neural network for evaluating the thermal insulation performance. The experimental data were randomly assigned to the training and testing samples in a ratio of 4∶1, and three artificial neural network models, namely BPNN, RBFNN and GRNN, were established respectively. Three evaluation criteria, namely the decision coefficient R2, mean absolute error MAE and mean square error MSE, were proposed. The predicted value of holding time and the specific value of evaluation criteria were obtained by the algorithm, and the neural network with the best performance was optimized by the random walk algorithm. By comparing the three evaluation criteria of R2, MAE and MSE and the actual and predicted values of holding time, it was concluded that RBFNN neural network had the best performance, the most accurate precision and the best fitting. Its R2 was much higher than those of GRNN and BPNN neural networks, and MSE and MAE were much lower than those of GRNN and BPNN neural networks. The three evaluation indicators reached 0.999 93, 0.009 63 and 0.062 86, respectively. The performance of the optimized Random-Walk-RBFNN was further improved, R2 was increased by 0.004%, and MSE, MAE and running time were decreased by 60.02%, 34.20% and 5.29%, respectively. RBFNN neural network is the most outstanding in all aspects, which is more suitable for evaluating the thermal insulation performance of cold chain incubator. The optimized Random-Walk- RBFNN has better performance, improves R2, reduces MSE, MAE and running time, and achieves better evaluation performance.