人工神经网络在气井管理及动态预测中的应用

鄂尔多斯盆地苏里格气田东区管理的井数多,稳产难度大,这对气藏开发及管理提出了更高的要求,而基于人工神经网络的气井动态预测方法不需要提供具体的地质参数,也不完全依赖气井生产动态数据。因此具有独特的优点。预测可分为3个步骤:1建立气井生产动态特征指标,用几个特征数据简单地识别气井生产能力,迅速准确地把待识别的单井生产日报数据与知识库样本中的单井做出类比,找到最相似的气井;2把ARPS递减曲线参数(初始产量、初始递减率)和开发指标(初始压力、动储量)作为预测目标,并与动态特征识别指标一起组合成向量,形成ARPS递减预测人工神经网络的预测模型,并计算气井每个月的动态数据;3建立气井动态人工神经网络训练的知识库,为更准确地预测气井合理开发指标提供依据。该方法在苏里格气田东区得到了成功应用,在投产时间较长的一批井中,通过动态分析以及气藏数值模拟得到143口井的气井动态ARPS递减参数人工神经网络预测模型,组成了人工神经网络的训练样本(知识库),为更准确地预测气井合理开发指标提供了依据。

Application of artificial neural network to gas well management and performance prediction

It is a challenge to develop and manage the reservoirs in the eastern block of the Sulige Gas Field, Ordos Basin, due to a great number of wells and the difficulty of maintaining stable production. The gas well performance prediction method based on artificial neural network neither requiring specific geologic parameters nor entirely depending on gas well production performance data has unique advantages. Performance prediction includes three steps: (1) To establish characteristics indicators of gas well production performance, identify the production capacity of a gas well with a few characteristic indicators, rapidly make an accurate analogy between the daily reported production data of a single well to be identified and that of the sample single well in a knowledge base to find out the most similar wells; (2) To use ARPS decline curve parameters (initial yield and initial decline rate) and development indicators (initial pressure and dynamic reserve) as prediction objectives, together with performance characteristics identification indicators as vectors to form the prediction model of ARPS decline prediction artificial neutral network, and then calculate the monthly performance data of a gas well; (3) To establish a knowledge base for the training related to a gas well performance artificial neural network to provide a basis for the more accurate prediction of gas well rational development indicators. This method has been successfully applied in the eastern block of the Sulige Gas Field. Performance analysis and gas reservoir and numerical simulation were conducted for a number of wells which had been put into production for a long time to obtain the prediction models (i.e., artificial neutral network) of ARPS decline parameters of 143 gas wells. These models were collected as artificial network training samples stored in a knowledge base to provide a basis for the more accurate prediction of rational development indicators of a gas well.