Forecasting Performance of Production Wells in the Maibarara Geothermal Field using Artificial Intelligence

Abstract

Reservoir modeling is the current industry practice used to forecast the response of the geothermal resource to production and in predicting the power plant’s future generation. However, building and running numerical reservoir model is costly and time-consuming. Due to these limitations, Artificial Intelligence techniques are being explored as an effective alternative using their strong learning capabilities and computational efficiency. This study aims to predict future production trends using artificial intelligence and the field’s historical data. The two machine learning models, multivariate linear regression (MLR) and multilayer perceptron (MLP), were built and trained for three Maibarara production wells. The model performance was compared using two error metrics, root mean squared error and coefficient of determination score. This study showed that MLP performs better than MLR using 80% training data and 20% testing data. The forecasts using the MLP model showed decreasing steam flow trends of varying magnitude for each well. Results also showed that model performance is influenced by the training-and-testing ratio, model type, and data imbalance, e.g. data size and distribution. For further evaluation, it is recommended to acquire more data for improvement in long-term forecasting accuracy, incorporate other data types, e.g. temperature-related data, enthalpy, and tracer test data, and explore and integrate other AI models, such as recurrent neural networks and long short-term memory.