A large amount of data is accumulating during the exploitation of oil wells. The data characterize the operating mode and properties of the extracted raw materials. It is not always use in a systematic and objective way, and not all possibilities for their application have been explored. The work is aimed at obtaining an understanding of the possible use of an array of such data to analyze the state of the well and predict the timing when an accident may occur. Relevant data were selected and a comparative analysis of operating parameters before failures and parameters in normal operating modes was carried out. There is a correlation between the characteristics of the well operating mode and the probability of failures (in particular, due to production casing leaks, etc.). The output results of machine learning algorithms for the separation of emergency and normal operating states were analyzed. It is shown how the trained algorithms work on the entire period of well operation (presented in the data and excluded from training). A typical picture of daily forecasts of production casing leaks type pre-emergency states on wells where such failures were occurred is very different from normal operating wells. There are a series of positive predictions over long intervals until a production casing leak is detected. The article proposes an evaluation of the results at different time intervals and a possible interpretation for use in production. Many of the other failures intersect or overlap each other, which makes it difficult to perform a multi-class separation and unambiguous conclusions about the effectiveness of their prediction. The presented results, at least in part, can clarify the issue of the probability and timing of failures and be used in the oil production monitoring.
References
1. Salehi S., Hareland G., Khademi K. et al., Casing collapse risk assessment and depth prediction with a neural network system approach, Journal of Petroleum Science and Engineering, 2009, V. 69, no. 1–2, pp. 156–162, https://doi.org/10.1016/J.PETROL.2009.08.011
2. Song X., Liu Y., Xue L. et al., Time-series well performance prediction based on Long Short-Term Memory (LSTM) neural network model, Journal of Petroleum Science and Engineering, 2020, V. 186, https://doi.org/10.1016/j.petrol.2019.106682
