Deep-learning models have been employed for production forecasting in oil and gas reservoirs, but they often assume that each well operates independently, neglecting the connectivity and dynamic interactions between wells. This simplification can significantly compromise prediction accuracy. Therefore, graph convolutional networks (GCNs) have been applied to incorporate data from neighbouring wells. However, existing spatial-temporal GCN (ST-GCN) methods are mainly used for autoregressive tasks and face limitations in predicting newly developed wells and fully utilizing temporal neighbour interactions. This study introduces an ST-graph- level feature embedding (ST-GFE) method that provides accurate production forecasting for newly developed wells. It enhances forecasting by aggregating the historical data from neighbouring wells into a single feature vector. This aggregated vector, merging local and contextual information, contains richer information about the studied region.
We evaluate ST-GFE using a dataset of 6,605 Montney shale gas wells, incorporating formation properties, fracture parameters, and production history. The ST-GFE is integrated with a non-autoregressive encoder-decoder structure to do production forecasting. The findings demonstrate that ST-GFE significantly improves prediction accuracy for newly developed wells compared to the purely temporal models, such as recurrent neural network (RNN)-based and Transformer models. ST-GFE adapts to production changes in adjacent wells, providing accurate predictions across various application scenarios, including shut-in and in-fill drilling activities. Additionally, while traditional GCNs require a full-batch training approach that leads to scalability issues, the ST-GFE model treats each well and its surrounding wells as a graph, enabling batch training and significantly reducing memory usage. Furthermore, the model dynamically updates its forecasts with real-time production data, enhancing precision and relevance. Experimental results confirm that ST-GFE effectively leverages spatio-temporal dynamics and interactions between adjacent wells, further improving production forecasting accuracy. This method enhances predictions and generalization capabilities for new developing locations, broadening its applicability to various drilling and production scenarios.
