Solvent swelling and catalyst have significant effects on the volatile release behavior during coal pyrolysis process, not only affecting the total amount, but also affecting the release trend when the temperature rises. An accurate quantification of this relationship can promote the efficient utilization of pyrolysis volatiles. This work adopts machine learning (ML) methods to construct prediction models for four primary volatiles (CO, CH4, aromatic and aliphatic hydrocarbons) during the pyrolysis of coal samples treated by different metal loading and solvent swelling, to predict the instantaneous volatile content at any time in the pyrolysis process. The deep learning models constructed by Recurrent neural networks (RNN) were compared with traditional machine learning methods of artificial neuron network (ANN), random forest (RF) and support vector machine (SVM), and realized an obvious preponderance on prediction performance for hole release curve. The best long short-term memory (LSTM) models achieve the average R2 of 0.978, 0.900, 0.950 and 0.936 for the release curve prediction of CO, CH4, aromatic and aliphatic volatile. Further adjustments shown that the reduction of sampling points on curves has little effect on the model’s performance, which makes it feasible to condense the models and decrease the dimensions of input. An easy signal processing of integral transformation had an obvious improvement on the predictive power of models, with a highest increase of R2 with 0.14 for aromatics predicting. All of these results provide a good method for real-time tracking of the conversion process.