This study aims to leverage deep learning (DL) to forecast oil production dynamics, focusing on the transition from conventional to unconventional reservoirs in the Shengli Oilfield, China. Due to the discrepancy in geological characteristics and development methods, production data from different reservoir types exhibit heterogeneous data distributions, which impedes the forecasting performance of DL models. To address this challenge, this study proposes an adversarial transfer-assisted transformer model, which utilizes advanced transfer learning techniques to achieve more accurate and universally applicable production forecasting across different reservoir types.
Our method involves segregating production data from different reservoir types into distinct domains. We employ a joint transformer model to project these domains into a unified representation space. Transformer's multi-head attention mechanism is adept at capturing the long-term dependency in production dynamics. Subsequently, we implement an adversarial training strategy using the generative adversarial network (GAN) framework. With this strategy, the model will converge to a Nash equilibrium, wherein the discriminator cannot distinguish between different domains. This indicates that the representations of different domains have similar probability distributions, thereby enabling the model to achieve cross-reservoir production forecasting.
The effectiveness of our model is validated using production data from multiple blocks within the Shengli oilfield, including conventional fault-block reservoirs and unconventional tight reservoirs. The results are twofold: (1) The Transformer model, serving as the feature extractor, outperforms TCN, LSTM, and GRU in capturing the long-term decline trends of production curves. (2) The introduction of transfer learning reduces the forecast error by 31.3% for unconventional reservoirs, compared to existing DL methods that ignore data discrepancies between reservoir types. Furthermore, the visualization of intermediate features intuitively verifies our model's capability to align feature distributions of different domains.
This study devises a novel adversarial transfer-assisted transformer model that can enhance the generalization performance of production forecasting across diverse reservoir types. Its advantage lies in aiding the management of unconventional reservoirs, which are often limited by historical data scarcity. By bridging different reservoir types using domain-invariant features, our model can efficiently leverage knowledge from existing conventional reservoirs to facilitate unconventional tasks. This is of great significance for oilfields characterized by a variety of reservoir types, offering new insights for improving the accuracy and efficiency of production forecasting.
