Abstract: In view of the problem that the traditional time series model cannot predict the gas yield of natural gas wells with a short production cycle, the affinity propagation algorithm based on machine learning is used to perform unsupervised clustering for natural gas wells and divide the well group. Combined with the geological and engineering parameters of the well group, principal component analysis is carried out to capture the key factors affecting the fluctuation of gas yield. Then, the maximum likelihood estimation method is used to solve the category of well group that the gas well belongs to, and the convolution neural network of training time for the production data of clustering center for the required category is used to predict the gas yield of the natural gas well in the short term. The results show that the yield prediction model of gas well based on machine learning has an average prediction error of 5.53%, and if compared with the long and short-term memory network (with an average error of 8.98%) and gated cyclic network (with an average error of 9.06%), the prediction error is smaller, indicating that this model can be applied for the yield prediction of carbonatite gas well with a relatively short development cycle. The research results are of great significance for the application research of machine learning in oil and gas reservoir development.

Key words: yield prediction of gas well, deep learning, convolution neural network, carbonatite gas well, unsupervised clustering