Multi-task ensemble learning for real-time efficiency prediction and condition diagnosis in beam-pumping well systems

Currently, real-time efficiency prediction and condition diagnosis for beam-pumping well systems predominantly rely on single-task models. However, conventional single-task approaches typically require constructing two separate models, resulting in doubled computational resource demands and an inability to effectively exploit the coupling relationships between prediction and diagnosis tasks. To address these limitations, this paper proposes a series of multi-task learning-based models for real-time efficiency prediction and condition diagnosis of beam-pumping well systems. First, we propose a multi-task learning model based on multimodal dual-path fusion, which enables effective feature sharing between efficiency prediction and condition diagnosis tasks using a dual-path network architecture. Second, we introduce a multimodal cross-scale fusion multi-task learning model designed to effectively capture the coupling relationships among multimodal features across different scales. Finally, to further improve the prediction accuracy of system efficiency and enhance the precision of condition diagnosis, we develop a meta-learning-based dynamic voting ensemble multi-task learning model, which dynamically integrates predictions from multiple tasks. To verify the effectiveness of the proposed methods, comprehensive experiments and ablation studies are conducted on a dataset consisting of 3938 actual beam-pumping wells. Experimental results demonstrate that, compared with the baseline networks, the proposed method achieves significant performance improvements: relative to the comparison networks, it reduces RMSE by 18.91%–44.73%, increases R² by 13.03%–101.13%, and enhances Accuracy and F1-score by 2.20%–32.84% and 2.16%–34.31%, respectively. Compared with the multimodal cross-scale multi-task learning network, it decreases RMSE by 6.86%–32.69%, raises R² by 3.66%~38.61%, and improves Accuracy and F1-score by 2.30%–173.35% and 2.31%–293.73%, respectively. Relative to the multi-task learning-based on meta-learning dynamic voting ensemble, it reduces RMSE by 17.30%–36.80%, increases R² by 5.40%–20.00%, and boosts Accuracy and F1-score by 1.77%–5.92% and 1.76%–6.17%, respectively. When compared with the ablation networks, the baseline network achieves RMSE reductions of 6.05%–27.78%, R² improvements of 3.05%–25.51%, and Accuracy and F1-score enhancements of 4.59%–53.73% and 7.76%–65.18%, respectively. Furthermore, the network exhibits significant sensitivity to hyperparameters, with notable performance fluctuations observed under different hyperparameter configurations.