Abnormal dissolved gas radon concentrations are reliable seismic anomaly. Studying potential patterns from historical data to predict future trend helps to quickly detect anomaly, which is vital for research on seismo-radon mechanism. Dissolved gas radon concentration observations are typical time series data. Traditional methods to solve this problem are mainly based on linear algorithms, such as Auto-Regressive (AR) model and Auto-Regressive and Moving Average (ARMA) model, which are of limit fitting ability to make precise prediction. This paper adopted the state of the art deep learning technique, Long Short-Term Memory (LSTM), to model dissolved gas radon concentration. In our experiment, we selected 4 years of dissolved gas radon concentration data collected from Guzan seismo station. We selected 90% of the data to train our model and 10% of the data to test our model. Comparison experiments were carried out on the same dataset with Auto-Regressive (AR) model and Auto-Regressive and Moving Average (ARMA) Model. Results showed that LSTM gains smaller RSME than AR and ARMA, which means that LSTM gain higher precision than AR and ARMA in our dataset.