Perpustakaan Digital ITB

Production forecasting is essential for production management. Therefore, methods such as decline curve analysis (DCA) or reservoir simulation is commonly used for predicting production performance. However, decline curve analysis has become less reliable because its ability based only on univariate analysis and reservoir simulation is time-consuming and has assumptions limiting what it should be used for. Artificial neural network (ANN) could be an alternative method as ANN could do multivariate forecasting based on production history data. This study aimed to develop a production performance prediction model for a given geothermal well in the Patuha field using three supervised learning methods (recurrent neural network (RNN), gate recurrent unit (GRU), and long short-term memory (LSTM), a neural network type model). Each model is trained with varying percentages of training data to determine the minimum data set is required for the model to give a decent forecasting result. Input data parameters such as steam rate, valve position, wellhead pressure, reservoir pressure, and temperature will be expected to be the observation data for predicting the production rate. The comparison results show that GRU has the lowest mean squared error (MSE) value (0.0048) and mean absolute percentage error (MAPE) value (3.05%), also has the highest R2 value (0.943). The minimum data set for the model to give decent results is 50% training and 50% validation. These three methods could be applied appropriately when considering the data set quality, how many parameters are used, and how well the data is processed.