Abstract: Currently, most underground wind temperature predictions use BP neural networks. But their prediction precision is affected by the number of learning samples and they are prone to falling into local optima. Elman neural networks have local memory capability, which improves the stability and dynamic adaptability of the network. However, there are still problems such as slow convergence speed and easy falling into local optima. In order to solve the above problems, the particle swarm optimization (PSO) algorithm is used to optimize the weights and thresholds of the Elman neural network. A bottom air temperature prediction model based on the PSO Elman neural network is established. The analysis shows that the relative humidity of the inlet and outlet wind, the surface inlet wind temperature, the surface atmospheric pressure, and the depth of the shaft are the main influencing factors of the bottom air temperature. Therefore, they are used as input data for the model, and the output data of the model is the bottom air temperature. The experimental results on the same sample dataset show that the Elman model converges at 90 iterations and the PSO Elman model converges at 41 iterations. It indicates that the PSO-Elman model converges faster. Compared with the BP neural network model, support vector regression (SVR) model, and Elman model, the prediction error of the PSO-Elman model is significantly reduced. The mean absolute error, mean square error (MSE), and mean absolute percentage error are 0.376 0 ℃, 0.278 3, and 1.95%, respectively. The determination coefficient R² is 0.992 4, which is very close to 1, indicating that the prediction model has good predictive performance. The verification results of the example show that the relative error range of the PSO-Elman model is −4.69%-1.27%, the absolute error range is −1.06-0.29 ℃, and the MSE is 0.26. The overall prediction precision can meet the actual needs of the underground.