Predicting the change trend of inlet valve temperature can provide important reference for the safety and reliability of equipment operation in valve cooling systems. For the problems of difficult feature extraction and low prediction accuracy of traditional methods, a Time2Vec neural network (T2VNN) model is put forward for predicting inlet valve temperature. This model firstly extracts the features of inlet valve temperature by Time2Vec, a time series representation learning method, and then capitalizes on the advantages of the TCN and bidirectional LSTM to achieve probabilistic prediction by quantile regression. Finally, comparative experiments with different time steps and quantiles on several typical models are designed. The experimental results verify that the T2VNN model has higher prediction performance, and the effectiveness of each component in the model is demonstrated by ablation experiments.