Shale Gas Production Prediction Based on Modified Graph Attention and Memory-Augmented Neural Networks

Summary Predicting shale gas production is crucial for optimizing shale gas well operations. However, production is influenced by numerous factors such as reservoir conditions, engineering parameters, human interventions, and interwell interference, posing significant challenges to accurate prediction. To address these complexities, we present a hybrid neural network model that combines a modified graph attention network (MGAT) with a memory-augmented neural network (MANN). MGAT enhances the extraction of spatial features from shale gas wells, while MANN uses gated recurrent units (GRUs) as its core architecture. It further incorporates production-related factors within its external memory module. MANN combines these components through an attention mechanism (AM), enabling it to efficiently extract relevant information from the external memory, thereby enhancing predictive performance. The effectiveness of the model is validated using field data from six wells with spatial relationships, and its performance is benchmarked against traditional models. Model hyperparameters are optimized using multiobjective particle swarm optimization (MOPSO). The results indicate that the proposed model achieves superior accuracy and robust generalization capabilities. Additionally, an ablation study highlights the critical role of temporal and spatial correlations for optimal model performance.