Hybrid MSCNN-AttBiGRU neural network for time-series well performance forecast

Precise prediction of well performance is vital for making informed project decisions and promoting efficient, cost-effective development of oil and gas fields, notably in unconventional reservoirs. Machine learning (ML) has gained prominence as a robust alternative, leveraging its nonlinear mapping strengths over the traditional reservoir numerical simulation and decline curve analysis methods. However, deploying ML for time series production prediction with only one technology remains challenging owing to the intricate application scenarios and the high standards for prediction accuracy. In this paper, we propose a task-oriented hybrid framework named the MSCNN-AttBiGRU model for time-series production forecasting, which is designed to improve prediction performance by hybridizing the Multi-scale Convolutional Neural Network (MSCNN) and Attention-enhanced Bidirectional Gated Recurrent Unit (AttBiGRU). MSCNN is adopted to capture the multi-scale spatial features, ranging from coarse to fine details. AttBiGRU leverages the attention mechanism to assign varying weights to the hidden layer states of BiGRU at different time step values, thereby extracting key temporal patterns. Eight cases in total are conducted including four synthetic well cases with varying noise level and four actual field well cases with fluctuations. The results indicate that the MSCNN-AttBiGRU model achieves competitive performance among the evaluated methods by considering the critical multi-scale spatiotemporal features of production sequences. The proposed model has the capacity to serve as a useful candidate tool for accurate and robust time series production prediction and stimulation optimization design.