Sucker-rod pumping wells are the most widely used producing wells in China. 94% of the 200,000 oil wells in CNPC are sucker-rod pumping wells. It is urgent to reduce the cost of every single well based on the well diagnosis and optimization methods under the background of low oil price and the IoT. Rich working experience of field engineers could help them diagnose some conspicuous abnormal well conditions by electrical power curves easily, but the scientific diagnosis methods have still not be established, and the potential of electrical power curves of the producing well is far from being fully tapped.
The aim of this work is to diagnose the working condition of the sucker-rod pumping wells both on and under the ground based on the data from electrical power curves by machine learning. The methods shaped by the learning of the electrical power curves from nearly 600 wells mainly separate into 3 steps. The first is the diagnosis of conspicuous abnormal well conditions such as motor belt burning, motor belt slippage, two phase electrics, upper rod break, lower rod break et al. The prediction experience was obtained from the statistical learning of the mean and variance values after we equally split the 600 power curve values into 10 sub-groups. The second is the diagnosis of complex abnormal well conditions such as abnormal mechanical sound, slight tube leak, severe tube leak, pump stuck et al based on the combination of statistics and template vs diagnosed sample analysis. The third is the diagnosis of pumping conditions characterized by the remarkable prediction ability via deep learning. A surface well condition database was established and the corresponding electrical power curves were marked in real time. Based on the CNN technology, the model could recognize different pump working conditions such as insufficient liquid, gas influence, traveling valve leak, standing valve leak et al very well.
The work has been applied in the oil fields of Jilin and Daqing. The method has been tested on nearly thousand hundreds of producing well utilizing sucker rod pumping system. The model demonstrates very high accuracy with almost 90% similarity to the result diagnosed by corresponding pump dynamometers for large sample and 94% of abnormal well working conditions for small sample. What’s more, the work would reduce millions of investment on the sensors, equipment and manpower for the management of producing wells in CNPC each year in the context of industrial IoT.