Accurate forecasting of oil production is essential for optimizing resource management and minimizing operational risks in the energy sector. Traditional time-series forecasting techniques, despite their widespread application, often encounter difficulties in handling the complexities of oil production data, which is characterized by non-linear patterns, skewed distributions, and the presence of outliers. To overcome these limitations, deep learning methods have emerged as more robust alternatives. However, while deep neural networks offer improved accuracy, they demand substantial amounts of data for effective training. Conversely, shallow networks with fewer layers lack the capacity to model complex data distributions adequately. To address these challenges, this study introduces a novel hybrid model called Transfer LSTM to GRU (TLTG), which combines the strengths of deep and shallow networks using transfer learning. The TLTG model integrates Long Short-Term Memory (LSTM) networks and Gated Recurrent Units (GRU) to enhance predictive accuracy while maintaining computational efficiency. Gaussian transformation is applied to the input data to reduce outliers and skewness, creating a more normal-like distribution. The proposed approach is validated on datasets from various wells in the Tahe oil field, China. Experimental results highlight the superior performance of the TLTG model, achieving 100% accuracy and faster prediction times (200 s) compared to eight other approaches, demonstrating its effectiveness and efficiency.
@Article{cmc.2025.059869,
AUTHOR = {Dalal AL-Alimi, Mohammed A. A. Al-qaness, Robertas Damaševičius},
TITLE = {A Hybrid Transfer Learning Framework for Enhanced Oil Production Time Series Forecasting},
JOURNAL = {Computers, Materials \& Continua},
URL = {http://www.techscience.com/cmc/online/detail/22416},
ISSN = {1546-2226},
DOI = {10.32604/cmc.2025.059869}