Forecasting hourly oil well tubing pressure based on the Res-RL model

The prediction of key parameters is a challenging task for the Internet of Things (IoT) system in intelligent oilfields. We proposed a novel design concept that integrates hybrid hidden layers and hybrid residuals, and developed the Res-RL model for the mid-term prediction of tubing pressure. The results of model performance comparison experiments based on the actual tubing pressure time series of 15 oil wells show that the six prediction error indicators (MSE, MAPE, RMSE, MAE, single-step MAPE, standard deviation of single-step APE) of our proposed model are significantly smaller than those of hybrid neural networks with cascaded modules (CNN-LSTM, CNN-LSTM-SA), traditional recurrent neural network models (RNN, LSTM, GRU), and traditional machine learning models (SVR, XGBoost). The performance test results on the edge computing platform indicate that the inference time of our Res-RL model is less than 1 s, and the prediction error is similar to that on the computer platform. These results suggest that our model can provide pressure data prediction services for future oilfield equipment maintenance and intelligent management in IoT systems, and that our design concept is an effective way to design hybrid neural network models.