The first step of our research was to identify the variables involved in the petroleum production prediction task. The first modeling effort included both production time series and geoscience parameters as input variables for the model. Eight factors that influence production were identified; however, since only data for the three parameters of permeability, porosity, and first shut-in pressure were available, the three parameters were included in the model. The production rates of the three months prior to the target prediction month were also included as input variables. The number of hidden units was determined by trial and error. After training the neural network, a sensitivity test was conducted to measure the impact of each input variable on the output. The results showed that all the geoscience variables had limited (less than 5%) influence on the production prediction.
Therefore, the second modeling effort relied on a model that consists of time series data only. The training and testing error was only slightly different from those of the first model. Hence, we concluded that it is reasonable to omit the geoscience variables from our model. More details on the modeling efforts can be found in (Nguyen et al., 2004).