A Bayesian-optimized convolutional neural network bidirectional gated recurrent unit model for dynamometer card reconstruction in beam pumping units

Abstract. The surface dynamometer card (DC) of beam pumping units plays a critical role in monitoring well conditions and adjusting pumping parameters. While several data-driven approaches based on deep learning have been applied to this task, they often suffer from limitations in input data completeness, model tuning, and the effective integration of deep learning structures. To address these issues, this paper proposes a novel inversion model named BO-CNN-BiGRU, which combines convolutional neural networks (CNNs), bidirectional gated recurrent units (BiGRUs), and Bayesian optimization (BO). The model inputs are determined based on the mechanism of beam pumping systems. CNNs are used to extract deep spatial features from electrical signals, BiGRU captures complex temporal dependencies to reconstruct polished-rod load curves, and BO enables automatic hyperparameter optimization. Field experiments conducted using real oil well production data demonstrate that the BO-CNN-BiGRU model achieves superior accuracy, robustness, and practicality. The proposed method can provide more reliable indicator diagrams to support intelligent regulation and diagnosis of beam pumping operations.