A group of Russian scientists from Gazprom VNIIGAZ, National Research University Higher School of Economics and Engineering Technologies Center developed a new model of forecasting the gas production, which can significantly speed up the calculations and facilitates the gas wells performance evaluation. This development will help oil and gas companies to make faster decisions on the optimal way to manage the production, on selecting the optimal operations modes and on efficient planning the oil and gas fields development.
Modern oil and gas companies are widely using hydro-dynamic models to simulate gas fluctuations in the reservoirs and to receive rather accurate forecasts. However, such models require big data sets, significant calculating resources and time, which makes them poorly suited for rapid analysis or for quick evaluation of numerous scenarios, especially when dealing with a multibranch well network.
To simplify and speed up the calculations, Russian scientists proposed to replace complicated physical-mathematical models with an easier one capable of learning based on the already available data. Such proxy model allows for forecasting well performance without labor-intensive simulations. In this model, the wells are viewed as elements of the graph: each well is presented as a node, and the links between them show the strength of their mutual impact depending on the distance and geological conditions. Hence, the model takes into account not only specific features of separate wells, but their mutual impact as well.
The researchers used time series as input data: bottomhole pressure, its change vs the preceding step and the duration of the well’s operation. The mutual impact of the wells depending on the distance was calculated separately. After that the data were sent through several neural network blocks. Recurrent and ultra-precision networks analyzed the changes with time, and the positional connections of the wells were processes with the help of graph convolutions. After combining the results, the model developed the forecast of production profiles for each well at each moment of time.
The model was learning based on big data including both the numerical simulations and real measurements received from gas fields. Dozens of thousands of various scenarios were used covering a broad range of geological and technological conditions. To improve the forecasts quality, all the data were sent through denoising and antialiasing with specialized filtering techniques.
Th trials showed that the developed model provides for high accuracy given significantly lower time vs the traditional methods. Working with synthetic data, the average forecasting error did not exceed 2%, and working with real field data it made about 10%, which is considered acceptable for practical implications. The calculations were performed four times faster vs the classic hydro-dynamic models.
The researchers believe their model has the potential of becoming instrumental for oil and gas companies – both for online wells performance evaluation and for long-term gas fields development planning.
