Production Forecasting Based on Attribute-Augmented Spatiotemporal Graph Convolutional Network for a Typical Carbonate Reservoir in the Middle East

Meng Gao, Chenji Wei, Xiangguo Zhao, Ruijie Huang, Jian Yang, Baozhu LiEnergies 2022 · DOI

TL;DR

This paper introduces an Attribute-Augmented Spatiotemporal Graph Convolutional Network (AST-GCN) for improved oil production forecasting in carbonate reservoirs. By integrating producer production data, neighboring injector gas injection rates (GIR), and spatial well relationships, the model significantly outperforms traditional methods like RNS and GRU, demonstrating enhanced accuracy and robustness.

What Problem Does This Paper Address?

Accurate oil production forecasting is crucial for effective reservoir management and development planning. Traditional methods like Decline Curve Analysis (DCA) and Reservoir Numerical Simulation (RNS) have limitations. DCA is simplistic and human-dependent, while RNS is data-intensive, time-consuming, and requires extensive geological expertise. While deep learning models like GRU have shown promise, they often focus on single-well predictions and struggle to capture complex inter-well dynamics and the influence of external factors like gas injection from neighboring wells. This paper addresses the challenge of simultaneously incorporating spatial well patterns, historical production data, and the dynamic influence of gas injection from adjacent wells to achieve more accurate and robust multi-well production forecasts.

Methods

The core of this work is the proposed Attribute-Augmented Spatiotemporal Graph Convolutional Network (AST-GCN). This hybrid model combines two key components:

  • Spatiotemporal Graph Convolutional Network (GCN): This part leverages GCNs to capture the spatial relationships between wells. It models the well network as a graph, where wells are nodes and connections (e.g., proximity, influence) are edges. This allows the model to learn from the spatial dependencies and interactions between producers.
  • Gated Recurrent Unit (GRU): GRUs are used to model the temporal dependencies within the production data. They are effective at capturing long-term trends and overcoming issues like gradient vanishing/exploding common in simpler recurrent networks.

The "Attribute-Augmented" aspect is crucial. The model explicitly incorporates the Gas Injection Rate (GIR) from neighboring injectors as a dynamic attribute. This is achieved through an "attribute augmentation unit" that encodes and integrates this GIR information into the GCN framework. This allows the model to understand how gas injection by one well can influence the production of others, a factor often overlooked by single-well models. The model's workflow involves data collection, preprocessing (denoising, normalization, splitting), training the AST-GCN, and then evaluating its performance against established methods.

Datasets

The study utilizes real-world production and gas injection data from a carbonate reservoir in the Middle East. Specifically:

  • Production Data: Raw daily oil production history for five producers over 2242 days (February 1, 2011, to March 22, 2017).
  • Gas Injection Data (GIR): Gas injection rates from seven gas injectors.
  • Geological Data: Reservoir permeability, thickness, and well distances were used to derive producer-injector relationship coefficients.

The data was preprocessed, including denoising using a Kalman filter and normalization using Min-Max Scaling. The dataset was split into an 87% training set and a 13% testing set. No new benchmark datasets were introduced; the focus was on applying the proposed model to existing reservoir data.

Key Insights and Ideas

  • The Power of Context: The most significant insight is that production forecasting is not just about a well's own history. The influence of neighboring wells, particularly gas injectors, plays a critical role and can be effectively captured by deep learning models.
  • Attribute Augmentation is Key: Simply using production data is insufficient. Augmenting the model with dynamic attributes like GIR from adjacent injectors dramatically improves prediction accuracy.
  • Spatial-Temporal Synergy: Combining graph convolutional networks (for spatial relationships) with recurrent neural networks (for temporal dynamics) creates a powerful framework for understanding complex reservoir behavior.
  • Robustness Matters: The model's ability to maintain accuracy even when subjected to simulated noise highlights its reliability in real-world scenarios where data can be imperfect.

Key Results

The AST-GCN model demonstrated superior performance compared to traditional RNS and GRU models:

  • Single-Well Prediction:
    • AST-GCN achieved significantly lower errors: 63.2% lower MedAE, 37.3% lower MAE, and 16.1% lower RMSE compared to GRU.
  • Compared to RNS, AST-GCN showed even greater improvements: 62.9% lower MedAE, 44.6% lower MAE, and 28.9% lower RMSE.
  • Accuracy was also higher: 15.9% higher than GRU and 35.8% higher than RNS.
  • Well Pattern Prediction:
    • For predicting the production of all five producers simultaneously, AST-GCN outperformed RNS with 41.2% lower RMSE, 64.2% lower MAE, and 75.2% lower MedAE.
  • The accuracy for well pattern prediction was 29.3% higher than RNS.
  • Robustness: When Gaussian noise was added to the data, the average changes in AST-GCN's performance metrics were minimal (3.3% for MedAE, 0.4% for MAE, and 1.2% for RMSE), indicating strong robustness.

Limitations and Future Work

The authors acknowledge several limitations:

  • Data Scarcity: Due to limited data, important parameters like Bottom Hole Pressure (BHP) and Well Head Flow Pressure (WHFP) could not be included, which are known to significantly impact production.
  • Static Producer-Injector Coefficients: The producer-injector relationship coefficient was treated as a constant value derived from geological data. In reality, these relationships can be dynamic and influenced by various factors.
  • Future Work:
    • Improving the accuracy and real-time nature of splitting gas injection rates.
  • Incorporating more dynamic elements into the attribute-augment matrix, such as permeability and choke size.
  • Developing methods to quickly find optimal hyper-parameters for the model to improve efficiency.
  • Considering additional data like pressure data for even more comprehensive forecasting.
  • The authors also suggest that the model can be used for downward distribution to predict the economics of development management units and single-well reservoirs, providing a basis for enterprise management decision-making.

Summary updated

Abstract

Production forecasting plays an important role in development plans during the entire period of petroleum exploration and development. Artificial intelligence has been extensively investigated in recent years because of its capacity to extensively analyze and interpret complex data. With the emergence of spatiotemporal models that can integrate graph convolutional networks (GCN) and recurrent neural networks (RNN), it is now possible to achieve multi-well production prediction by considering the impact of interactions between producers and historical production data simultaneously. Moreover, an accurate prediction not only depends on historical production data but also on the influence of neighboring injectors’ historical gas injection rate (GIR). Therefore, based on the assumption that introducing GIR can enhance prediction accuracy, this paper proposes a deep learning-based hybrid production forecasting model that is aimed at considering both the spatiotemporal characteristics of producers and the GIR of neighboring injectors. Specifically, we integrated spatiotemporal characteristics and GIR into an attribute-augmented spatiotemporal graph convolutional network (AST-GCN) and gated recurrent units (GRU) neural network to extract intricate temporal correlations from historical data. The method proposed in this paper has been successfully applied in a well pattern (including five producers and seven gas injectors) in a low-permeability carbonate reservoir in the Middle East. In single well production forecasting, the error of AST-GCN is 63.2%, 37.3%, and 16.1% lower in MedAE, MAE, and RMSE compared with GRU and 62.9%, 44.6%, and 28.9% lower compared with RNS. Similarly, the accuracy of AST-GCN is 15.9% and 35.8% higher than GRU and RNS in single well prediction. In well-pattern production forecasting, the error of AST-GCN is 41.2%, 64.2%, and 75.2% lower in RMSE, MAE, and MedAE compared with RNS, while the accuracy of AST-GCN is 29.3% higher. After different degrees of Gaussian noise are added to the actual data, the average change in AST-GCN is 3.3%, 0.4%, and 1.2% in MedAE, MAE, and RMSE, which indicates the robustness of the proposed model. The results show that the proposed model can consider the production data, gas injection data, and spatial correlation at the same time, which performs well in oil production forecasts.
BibTeX
@inproceedings{gao2022production,
  author = {Meng Gao and Chenji Wei and Xiangguo Zhao and Ruijie Huang and Jian Yang and Baozhu Li},
  title = {{Production Forecasting Based on Attribute-Augmented Spatiotemporal Graph Convolutional Network for a Typical Carbonate Reservoir in the Middle East}},
  booktitle = {Energies},
  year = {2022},
  doi = {10.3390/en16010407},
}
Figure