Hybrid Data Driven Approach for Reservoir Production Forecast

Analytical models, in particular Decline Curve Analysis (DCA) are widely used in the oil and gas industry. However, they are often solely based on production data from the declining wells and do not leverage the other data available in the field e.g. petrophysics at well, completion length, distance to contacts... This paper describes a workflow to quickly build hybrid models for reservoir production forecast based on a mix of classic reservoir methods and machine learning algorithms. This workflow is composed of three main steps applied on a well by well basis. First, we build an object called forecaster which contains the subject matter knowledge. This forecaster can represent parametric functions trained on the well itself or more complex models that learn from a larger data set (production and petrophysics data, synthesis properties). Secondly this forecaster is tested on a subset of production history to qualify it. Finally, the full data set is used to forecast the production profile. It has been applied to all fluids (oil, water, gas, liquid) and revealed particularly useful for fields with large number of wells and long history, as an alternative to classical simulations when grid models are too complex or difficult to history match. Two use cases from conventional and unconventional fields will be presented in which this workflow helped quickly generate robust forecast for existing wells (declining or non-declining) and new wells. This workflow brings the technology, structure and measurability of Data Science to Reservoir Engineering. It enables the application of the state of the art data science methods to solve concrete reservoir engineering problems. In addition, forecast results can be confronted to historical data using what we call "Blind Testing" which allows a quantification of the forecast uncertainty and avoid biases. Finally, the automated workflow has been used to generate a range of possible realizations and allows the quantification the uncertainty associated with the models.