[URTeC 2023] How Much Data is Needed to Create Accurate PDP Forecasts?​ (ID 3870872)

Abstract:

Production forecasts are critical for determining well productivity, estimating future cash flow, asset valuations, and informing key business decisions. When using traditional decline curve analysis (DCA) or newer machine learning (ML) methods to forecast the production of existing wells, it is important to ensure there is sufficient production decline history for maximum accuracy and reliability. The question then becomes: how much data does each method require, and can one method make more accurate forecasts than the other with less data? In this study, we replicate a real-world scenario by hiding and re- introducing increments of production data to determine the minimum decline history needed to create accurate forecasts for DCA and ML methods. Three models were created, a “Limited ML model” using only production history as a feature, a “Full ML model” using a full feature set, and a DCA model. The Limited ML model allows us to analyze the effects of well-specific features on PDP forecasts. Our study shows that the Full ML model can forecast cumulative production with a median absolute percent error (MAPE) of 14.4% with only 90 days of decline data. This is a significant improvement over its Limited ML model counterpart, which had a MAPE of 15.6%. Comparison of the limited ML and DCA models showed similar performance, with the DCA model providing a MAPE of 16.1% with 90 days of decline data. Errors continue to decrease with all models converging at a MAPE of ~1% after 1080 days of decline data. In addition, program level analysis of model performance showed that the ML models had a peak aggregate percent error (APE) between 23% – 29%. This is a significant improvement over the DCA peak APE of 54%. Our study shows that the improvement in forecast accuracy with limited data is due to the ability of ML models to leverage additional information like geology, completions, and spacing to decline wells accurately. Furthermore, the rigidity of the curve-fit workflow DCA models use does not allow DCA forecasts to react appropriately to unexpected deviations from an ideal hyperbolic decline. Meanwhile, ML models can learn and react appropriately to change the wells’ forecast when provided with appropriate production