Hybrid weights structure model based on Lagrangian principle to handle big data challenges for identification of oil well production: A case study on the North Basra oilfield, Iraq

• Published in: Engineering applications of artificial intelligence 2024-12, Vol.138, p.109465, Article 109465
• Main Authors: Homod, Raad Z., Albahri, A.S., Munahi, Basil Sh, Alamoodi, A.H., Hussein, Ahmed Kadhim, Albahri, O.S., Alhasnawi, Bilal Naji, Al-Mudhafar, Watheq J., Mahdi, Jasim M., Yaseen, Zaher Mundher
• Format: Article
• Language: English
• Subjects: Big data identification; Grey box model; Hybrid weights structure; Lagrange's interpolation; Stochastic dynamical system
• ISSN: 0952-1976
• DOI: 10.1016/j.engappai.2024.109465

The identification of the oilfield production flow rate, which is a function of the wellhead pressure, where both are characterized as a complex, nonlinear stochastic dynamical system and heterogeneity phase coupling with a very high delay time. Hence, such a characterization of the system will not be able to fulfil the purpose of creating a conventional model, in addition, it needs the recruitment of a large dataset. The dataset is collected using the log reader agent on each oil well and is arranged in rows and columns where each column contains 16 million rows for each vector of the inputs. At this end, in order to handle such kind of task, hybrid weights (training weights and estimated weights) are combined to create the proposed Lagrange's interpolation model based on the hybrid weight structure (LIMBHWS) which is a type of grey box model. The LIMBHWS algorithm plays a crucial role in optimizing model outputs via nonlinear regression. Extracting odd-indexed elements from each dataset vector to use them as a training dataset effectively halves the required training time. Also, easily the LIMBHWS computes the estimated weight by interpolation methods for their analogues of training weights. The results of the proposed algorithm LIMBHWS show that 50% of training time is eliminated, where the mean absolute errors (MAE) are 8.976, 14.328 and 23.167 for the proposed model, training weights model and the model of the estimated weight respectively. •Combined two types of weights (training and estimating) to obtain the proposed model weights (hybrid weights).•Converting the discrete dataset into hybrid weight enables LIMBHWS to predict continuous outputs.•The sub-algorithm of estimated weight is used to tackle an extremely large dataset.•Used MLP structure strategy to reduce the computational training cost and rapidly output response.•The investigation of outputs for the proposed model shows significantly close to real values.