The decline curve is critical for understanding and optimizing the financial implications of well placement, pad design, and completions designs. However, the direct impact of engineering decisions on the parameters which govern the decline curve remains an open question. Here, we engage in a multi-basin machine learning approach, aimed at describing how operational decisions affect each Arps equation parameter, for each fluid phase in production. We find that completions designs have the largest effect on initial production, whereas interwell spacing is more influential for observed decline rates. The importance of geologic features tend to vary per fluid stream, with clay volume being particularly important for oil production, and water saturation for water.
The decline curve is one of the most important elements of the oil and gas industry. It is a tool that allows engineers to forecast oil, gas, and water production quickly and is critical for defining the rate of revenue and cash flow for upstream participants. Therefore, understanding and predicting its shape using the Arps framework can help producers constrain critical top and bottom line business metrics (Walsh and Lake 2003).
It is known that typical values for Arps equation parameters vary across basins and across fluid windows. There have also been studies that analyze differences to reservoir responses to completion designs, but these are typically described in terms of changes in EUR, or in terms of percent uplift / downlift at a given production day (e.g. Cross et al. 2020). The effect of completions designs – and their interactions with reservoir characteristics – on descriptions of decline curves is an open question with direct applications to financial planning, pad design, and cross-basin well analogues.
This study uses machine learning to explore which engineering and subsurface features/variables most influence each of the Arps parameter values. We accomplish this by training a multitarget tree-based ensemble to predict Arps equation parameters directly from completion design, pad design, and geology. We then evaluate model explanations for predictions using Shapley values to determine which features have the greatest positive or negative impact on each equation parameter. Secondarily, we are interested in assessing the generalizability of these results across basins, so we repeat the analysis in both the Midland and Williston basins where both the geology and development practices differ.