What question did this study set out to answer?

To develop and validate TimeSenseNet, a model that enhances accuracy and robustness in shale oil production forecasting using physics and data fusion.

February 26, 2026

TimeSenseNet: A Physics-Data Fusion Model for Shale Oil Production Forecasting

Key Points

To develop and validate TimeSenseNet, a model that enhances accuracy and robustness in shale oil production forecasting using physics and data fusion.
Model integrates Time2Vec time encoding, CNN, BiLSTM, and attention mechanisms.
Utilizes monthly production data from 18 wells in the Bakken Formation and 5 wells in the Eagle Ford Formation for validation.
Emphasizes the incorporation of physical reservoir parameters like lithology, porosity, and pressure for feature extraction.
Achieved an average R2 of 0.81 across validation wells in Eagle Ford, with peak performance at 0.91.
Lowest measured average MAE at 205.70, RMSE at 434.86, and NRMSE at 0.08, indicating strong predictive reliability.
Successfully allows real-time updates and long-term prediction capabilities, reducing error accumulation.

Abstract

To address the complex production characteristics of unconventional oil reservoirs, we propose TimeSenseNet, a prediction model designed to improve both the accuracy and the robustness of shale oil production forecasting. The model innovatively integrates Time2Vec time encoding, Convolutional Neural Networks (CNN), Bidirectional Long Short-Term Memory networks (BiLSTM), and a Transformer-based attention mechanism. This architecture automatically captures complex nonlinear patterns in time-series data and simplifies traditional preprocessing procedures, while explicitly incorporating physical reservoir parameters including lithology, measured depth (MD), true vertical depth (TVD), porosity, permeability, reservoir temperature, and pressure, to guide feature extraction and temporal modeling. Monthly production data from 18 wells in the Bakken Formation were used, with 80% for training and 20% for testing, and 5 wells in the Eagle Ford Formation, including both horizontal and vertical wells, were reserved for independent validation. Results on the Eagle Ford Formation validation set show that TimeSenseNet achieves an average R2 of 0.81 across all wells, with peak single-well performance reaching 0.91. It also attains the lowest average mean absolute error (MAE) (205.70), root-mean-square error (RMSE) (434.86), normalized root mean squared error (NRMSE) (0.08), and weighted absolute percentage error (WAPE) (0.21), demonstrating strong generalization to the unseen reservoirs. This study establishes a closed-loop process of data governance, model optimization, and field validation, showing that TimeSenseNet provides accurate short-term production forecasts and supports long-term rolling predictions with reduced error accumulation. The separation of static reservoir features and dynamic production data allows efficient real-time updates when new data become available. These capabilities demonstrate TimeSenseNet’s potential for data-driven decision-making and more efficient shale oil production in unconventional reservoirs.

Bookmark

TimeSenseNet: A Physics-Data Fusion Model for Shale Oil Production Forecasting

Key Points

Abstract

Cite This Study