Production prediction is crucial for the recovery of hydrocarbon resources. However, accurate and rapid production forecasting remains chall
Production prediction is crucial for the recovery of hydrocarbon resources. However, accurate and rapid production forecasting remains challenging for unconventional reservoirs due to the complexity of the percolation process and the scarcity of available data. To address this problem, a novel model combining a long short-term memory network (LSTM) and support vector regression (SVR) was proposed to forecast tight oil production. Three variables, the tubing head pressure, nozzle size, and water rate were utilized as the inputs of the presented machine-learning workflow to account for the influence of operational parameters. The time-series response of tight oil production was the output and was predicted by the optimized LSTM model. An SVR-based residual correction model was constructed and embedded with LSTM to increase the prediction accuracy. Case studies were carried out to verify the feasibility of the proposed method using data from two wells in the Ma-18 block of the Xinjiang oilfield. Decline curve analysis (DCA) methods, LSTM and artificial neural network (ANN) models were also applied in this study and compared with the LSTM-SVR model to prove its superiority. It was demonstrated that introducing residual correction with the newly proposed LSTM-SVR model can effectively improve prediction performance. The LSTM-SVR model of Well A produced the lowest prediction root mean square error (RMSE) of 5.42, while the RMSE of Arps, PLE Duong, ANN, and LSTM were 5.84, 6.65, 5.85, 8.16, and 7.70, respectively. The RMSE of Well B of LSTM-SVR model is 0.94, while the RMSE of ANN, and LSTM were 1.48, and 2.32.
