Region-Wide Production Forecasting Using Dynamic Model Selection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current basin forecasting technologies face challenges in generating accurate production forecasts for wells with limited or no production data, as they rely on existing data and struggle to predict future production effectively.
Innovation Solution
A method that identifies base data for wells, selects appropriate models such as rich machine learning, location-based machine learning, and decline curve models, and generates forecasts using these models to predict production individually for each well, then aggregates these forecasts to provide a region-wide forecast, utilizing different techniques based on data availability.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional basin forecasting technology relies on existing production data, then forecasting accuracy is improved for wells with sufficient data, but forecasting capability deteriorates for wells with little or no production data
Solution Approach 1:
The patent segments the basin into multiple regions based on geological characteristics, well density, and production patterns. Different forecasting models are applied to different regions: machine learning models for data-rich regions and decline curve models for data-poor regions. This segmentation allows the system to optimize forecasting accuracy for each region while maintaining overall basin-wide forecasting capability.
Solution Approach 2:
The patent introduces an intermediary hierarchical structure where regional forecasts serve as intermediaries between individual well data and basin-wide totals. The system first generates forecasts at the well level, then aggregates to regional level, and finally combines regions to produce basin-wide forecasts. This intermediary hierarchy allows information to flow efficiently through multiple levels, enabling accurate forecasting even when individual well data is limited.
2Adaptability or versatility
If multiple forecasting models are used to handle different data scenarios, then forecasting versatility is improved, but system complexity increases
Solution Approach 1:
The patent implements a dynamic model selection mechanism that automatically chooses the appropriate forecasting model based on data availability and quality for each well. The system assesses the amount and quality of production data for each well and dynamically selects between machine learning models, decline curve models, or hybrid approaches. This dynamic adaptation allows the system to maintain versatility while managing complexity through automated decision-making rules.
Solution Approach 2:
The patent applies the principle of local quality by using different forecasting models tailored to specific local conditions. Machine learning models with multiple features are applied where sufficient production and well data are available, while simpler decline curve models are used where data is limited. This localized model selection ensures each well receives the most appropriate forecasting approach for its specific data situation, optimizing overall system performance.
3Measurement precision
If comprehensive machine learning models with multiple features are used, then forecasting accuracy is improved, but computational resources increase
Solution Approach 1:
The patent applies partial action by using comprehensive machine learning models with multiple features only where necessary - specifically for wells with sufficient production data and in regions where data quality supports complex modeling. For wells with limited data, the system uses simpler decline curve models that require fewer computational resources. This selective application of complex modeling achieves high accuracy where needed while conserving computational resources overall.
Data Source
AI summary
A method for rapid region wide production forecasting includes identifying base data of a well in a plurality of wells of a region; selecting, using the base data and from a set of a models comprising a rich machine learning model, a location based machine learning model, and a decline curve model, a well model; and generating, based on the selecting, a forecasted production of the well using the base data and the well model. The method further includes aggregating a plurality of forecasted productions of the plurality of wells, the plurality of forecasted productions including the forecasted production, to generate a region forecast using the rich machine learning model, the location based machine learning model, and the decline curve model; and presenting the region forecast.