Scholarlink

<p>Electric submersible pumps (ESPs) are essential components of oilfield production systems, and are widely used across various operational environments. However, production prediction for ESPs is particularly challenging because of the influence of multiple factors, including geological conditions, reservoir properties, and equipment status. Current approaches rely primarily on physical mechanism models or conventional data-driven approaches, which often fail to effectively capture complex temporal relationships and multidimensional features, resulting in limited prediction accuracy. To address these challenges, this paper proposes an advanced LLM-TFNet model based on large language model (LLM) to accurately predict the daily output of ESPs. The proposed model introduces an input conversion layer, a patch reprogramming layer, and an output projection layer, and integrates multilevel modal fusion (MLMF) to reprogram the LLM. This effectively reframes the ESP time series prediction task as a language modeling task. Additionally, we introduce a time–frequency channel attention mechanism (TFCAM) based on the discrete cosine transform (DCT), enabling comprehensive modeling of combined features in both the time and frequency domains. This approach significantly enhances the ability to capture the intricate temporal patterns of ESPs. To improve learning efficiency, an adaptive expert allocation mechanism (AEAM) is introduced, which dynamically assigns networks to select suitable expert subnetworks for feature extraction to improve computational efficiency and prediction accuracy. Comparative experiments against machine learning models, transformer-based methods, and LLM-based methods demonstrate that our method achieves a performance gain of 1.07%–20.68% across various evaluation metrics. Ablation experiments further validate the effectiveness of our proposed framework. When real production data from an offshore oilfield were tested, the model achieved a prediction accuracy of 95% with an average error of less than 0.8 tons, meeting practical production requirements.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

LLM-TFNet: a method of production prediction on electric submersible pump wells based on large language model

  • 期刊论文

  • 发表日期：

    2025年10月9日

查看全文
  • Faming Gong,
  • Hao Wang,
  • Chengze Du,
  • Junjie Hao,
  • Jingcheng Gao

摘要

Electric submersible pumps (ESPs) are essential components of oilfield production systems, and are widely used across various operational environments. However, production prediction for ESPs is particularly challenging because of the influence of multiple factors, including geological conditions, reservoir properties, and equipment status. Current approaches rely primarily on physical mechanism models or conventional data-driven approaches, which often fail to effectively capture complex temporal relationships and multidimensional features, resulting in limited prediction accuracy. To address these challenges, this paper proposes an advanced LLM-TFNet model based on large language model (LLM) to accurately predict the daily output of ESPs. The proposed model introduces an input conversion layer, a patch reprogramming layer, and an output projection layer, and integrates multilevel modal fusion (MLMF) to reprogram the LLM. This effectively reframes the ESP time series prediction task as a language modeling task. Additionally, we introduce a time–frequency channel attention mechanism (TFCAM) based on the discrete cosine transform (DCT), enabling comprehensive modeling of combined features in both the time and frequency domains. This approach significantly enhances the ability to capture the intricate temporal patterns of ESPs. To improve learning efficiency, an adaptive expert allocation mechanism (AEAM) is introduced, which dynamically assigns networks to select suitable expert subnetworks for feature extraction to improve computational efficiency and prediction accuracy. Comparative experiments against machine learning models, transformer-based methods, and LLM-based methods demonstrate that our method achieves a performance gain of 1.07%–20.68% across various evaluation metrics. Ablation experiments further validate the effectiveness of our proposed framework. When real production data from an offshore oilfield were tested, the model achieved a prediction accuracy of 95% with an average error of less than 0.8 tons, meeting practical production requirements.