High-temperature alkali metal heat pipes are of significant application value in the cooling systems of space nuclear reactors due to their excellent heat transfer performance. However, the non-steady heat transfer characteristics during the frozen startup process of heat pipes are complex, and temperature response prediction has remained a challenging research issue. Traditional numerical modeling methods, although possessing good physical interpretability, have limited prediction accuracy and high computational costs under the complex conditions of the frozen startup phase. To address this, this paper proposes a prediction network based on Feedforward Neural Network (FNN), utilizing large-scale training data generated by the numerical model for network training. In the network design, early stopping callback is introduced to prevent overfitting and improve the fitting speed. Experimental results show that the model can accurately predict the temperature distribution state during the heat pipe's cold startup process, with predictions being largely consistent with the calculated results and closely matching the experimental data. Additionally, the method can accurately predict the startup process of heat pipes with different parameters, providing an efficient and accurate solution for predicting the complex heat transfer process during the frozen startup of high-temperature alkali metal heat pipes.