Abstract: To address the issues of high labor intensity, low testing frequency, and high testing costs associated with traditional manual testing methods for the dynamic liquid level in pumping unit wells, the article employs the Pearson correlation coefficient analysis method to investigate the correlation between 29 automatically collected characteristic parameters of pumping unit wells and the measured dynamic liquid level, ultimately identifying 13 key characteristic parameters. Utilizing machine learning techniques, including XGBoost, LightGBM, and BP neural network, distinct dynamic liquid level prediction models for pumping unit wells were developed. Through the input of 13 key characteristic parameters into these models, an evaluation of their prediction outcomes was conducted. The evaluation revealed that a singular prediction model was inadequate for all pumping unit wells. Consequently, a variable weight combination model, founded on the three prediction models, was formulated. Numerous field applications in Changqing Oilfield have demonstrated that, in comparison to traditional manual testing methods, this approach achieves an average relative error within 5%, a testing efficiency increase of over 150,000 times, a reduction in labor intensity by over 90%, a testing frequency increase exceeding 2,000 times, and a significant 96% reduction in testing costs. In conclusion, the variable weight combination dynamic liquid level prediction model effectively addresses the challenges posed by high labor intensity, low testing frequency, and high testing costs inherent to traditional manual testing methods, thereby offering novel insights for dynamic liquid level testing in domestic oil fields.