Dynamic Production Forecasting using Artificial Neural Networks customized to historical well Key Flow Indicators

The existing decline curve analysis (DCA) equations, some with valid theoretical justifications, cannot directly react to changes in operating conditions. Thus, they all assume constant operating conditions over the flowing life of a well. This however is an obvious oversimplification. This paper begins by briefly reviewing Gilbert's equation for flowrate prediction and then the C-curve and Logistic growth model DCA theories. The above review serves to identify well key flow indicators (KFI) and performance drivers. Subsequently, a forecasting approach which involves building artificial neural network (ANN) frameworks and training them on well KFI data is presented. Using trained ANNs, production forecasts were generated for three oil wells in the Niger-Delta producing from separate reservoirs under different flow regimes. The results were compared to forecasts from traditional DCA methods and material balance simulation, as well as with future production from the wells themselves. The results indicated that trained ANNs are capable of generating better performance curves than traditional DCA, with forecasts tying closely with results of material balance simulation and measured future well production rates. The ability of trained ANNs to evaluate the effect of changes in operating conditions (i.e. FTHP, GOR and water-cut) on production profiles and reserves drainable by wells, allows for scenario forecasting which is invaluable in field development planning. This is illustrated with field cases. This paper also presents a novel approach to evaluating the optimal hyperparameter configuration (i.e. the number of layers, neuron count per layer, dropout, batch size and the learning rate) required to minimize the loss function whilst training an ANN on any given dataset. This should prove invaluable to engineers and geoscientists integrating deep learning into sub-surface analyses.