A Model Embedded with Development Patterns for Oilfield Production Forecasting

Machine learning models that only use data for training and forecasting oilfield production have a sense of disconnection from the physical background, while embedding development patterns in them can enhance interpretability and even improve accuracy. In this paper, a novel multi-well production forecasting model embedded with decline curve analysis (DCA) is proposed, enabling the machine learning model to incorporate physical information. Moreover, an improved particle swarm optimization algorithm is proposed to optimize the hyperparameters in the loss function of the model. These hyperparameters determine the importance of the overall DCA and each module in training, which traditionally requires expert knowledge to determine. Simulation results based on the benchmark reservoir model show that the model has better forecasting ability and generalization performance compared to typical machine learning methods.