The accuracy of predicting oilfield production via machine learning algorithms is closely related to the representativeness and quantity of training samples. Generally, oilfield production data or oil well production data are used to construct training samples. There is the "small sample" problem when the oilfield is used for the training samples. To use oil wells as training samples, it is difficult and time-consuming to manually mark training samples that can represent the oilfield production decline, because old oil fields generally have many development layers, long production history and many production batches of oil wells. The production data of oilfield and oil well production data are organically integrated to construct training samples, and the production intelligent prediction model is established to predict the production of the oilfield. Firstly, the K-means clustering algorithm of unsupervised learning is used to perform cluster analysis on oil wells based on effective thickness, porosity, permeability, saturation and other observed values, identify the production decline category, and convert each type of oil well into a typical oil well as a representative of this type of oil well. Secondly, typical wells are taken as prediction objects, and the number of training samples is increased by randomly extracting wells proportionally from each type of well, that is, the production data of typical wells and wells are fused to construct training samples. Finally, a model is built based on LSTM neural network to predict the production of typical wells, and then predict the oilfield production. The research results show that this method not only solves the "small sample" problem of oilfield data as training samples, but also reduces the difficulty and workload of labeling oil well data as training samples, and the accuracy meets the requirements of field production, which has certain guiding significance for the engineering application of intelligent prediction of oil and gas production.