Perpustakaan Digital ITB

Reliable steam production forecasting is vital for managing geothermal fields, yet conventional Arps decline curve analysis assumes a fixed decline in time and cannot use the wellhead pressure that governs steam deliverability. This study evaluates whether machine learning can forecast steam production more accurately than the Arps approach for a single dry steam well, Well-1. From 438 weekly records (2004 to 2012) split chronologically into 306 training, 66 validation, and 66 testing, three Arps models were fitted from the production peak to the pressure-normalized rate, while SVR, RF, and XGBoost were applied in a residual-hybrid scheme using the static and flowing wellhead pressures, with the annual decline rate as the primary, apple-to-apple comparison. The results agree: the best Arps model (exponential) gives an effective decline of about 4.268%/year, close to the machine learning declines of 4.579%/year (XGBoost), 4.578%/year (RF), and 4.875%/year (SVR), and over the test zone the machine learning models reproduce the measured rate closely (R2 of 0.983, 0.980, and 0.964), though these are hindcasts likely optimistic because the weekly series was interpolated from digitized data. It is concluded that machine learning complements rather than replaces conventional decline analysis; projected forward at about 4.579%/year, the steam rate falls from about 29.19 kg/s in 2012 to about 7.39 kg/s in 2042. It is recommended that future work test more wells, raw field data, and an independent wellhead-pressure forecast for genuine forward prediction.