Abstract: as load forecasting is crucial to the optimal dispatch of gas resources. In addition to the temporal characteristics such as trend and periodicity, there are also spatial characteristics between the load data and temperature and humidity data of neighboring gas regulator stations, which makes it difficult to model the gas load prediction mechanism and the model prediction accuracy is low. In response to the above problems, a gas load forecasting model based on the combination of autoregressive moving average model (ARIMA) and convolutional long-short-term neural network (ConvLSTM) is proposed. Firstly, the Pearson correlation coefficient is selected to analyze the gas load data and the variables strongly correlated with the gas load are selected as the input of the model. Secondly, the ARIMA model is used to remove the trend of the data to make it smooth, the ConvLSTM model is used to extract the spatio-temporal features in the data, and the parameters of the ARIMA-ConvLSTM model are optimized. Finally, the model is trained and verified through gas load data. Experimental results show that the prediction accuracy of the ARIMA-ConvLSTM model is 98.65%, which is better than the ARIMA model, ConvLSTM model and CNN-LSTM parallel combined model in terms of root mean square error, average absolute error, and absolute error percentage.