EMD-based multi-algorithm combination model of variable weights for oil well production forecast. (November 2022)

• Record Type: Journal Article
• Authors: Cao, Yu; Liu, Shanke; Cao, Xiaopeng; Liu, Xinyi; Hu, Huifang; Zhang, Tingting; Yu, Lijun
• Is Part Of: Energy reports. Volume 8(2022)
• Page Start: 13389
• Page End: 13398
• Publication Date: 2022-11
• Subjects: Petroleum engineering -- Petroleum production forecast -- Machine learning -- Combination algorithm -- Empirical mode decomposition; Power resources -- Periodicals; Energy industries -- Periodicals; Power resources; Periodicals; Electronic journals; 621.04205
• DOI: 10.1016/j.egyr.2022.09.140
• Languages: English
• ISSNs: 2352-4847

Abstract

Abstract: Oilfields in high or ultra-high water-cut period are of high nonlinearity and heterogeneity, thus complicated in its internal physics mechanism. Corresponding production data is often of small amount due to large timesteps. In addition, numerous mechanism factors that are usually a prerequisite for traditional data-driven prediction methods are to be determined with extra human effort on a well basis, deviating from the intention of oilfield digital transformation. In response to these obstacles, this research proposes an EMD (Empirical Mode Decomposition)-based variable-weight combination model algorithm (EMD-Combined model): It does not require inputs other than its own historical oil production time series, but rather automatically constructs and filters input features via EMD, time-lag reconstruction and Spearman correlation analysis instead. A dynamic Variable-Weight Combination Methods of four machine learning models with different strengths is constructed to carry out prediction. Experimental results show that the EMD-Combined model is more accurate than the solely use of any of the four single models listed above, or models solely exerted with EMD without model-wise combination. In terms of the ease of usage, deployment and "stability" of prediction quality, EMD-Combined model offers a possibility of improvement in real-life practice. Highlights: A novel time-varying variable-weight hybrid machine-learning model for petroleum well production forecast. Reduction of prediction complexity and feature engineering via input preprocessing by Empirical Mode Decomposition. Reduction of required input to only the historical time-series data of the prediction object. Automated aggregation of multiple prediction algorithms with different strengths to improve generalization across periods and between wells.