Forecasting Oil Production Time Series with a Population-Based Simulated Annealing Method

This paper addresses the oil production forecasting problem, closely related to the estimation of reserves remaining in oil fields. We present a new method, named SAM-oil, which combines simulated annealing metaheuristic with evolutionary computing techniques, statistical analysis and quality assessment of solutions. The method implements a learning-based scheme for pre-processing, modeling and forecasting of monthly oil production time series. Accuracy of point forecasts is compared between SAM-oil and a typical technique in petroleum engineering, known as decline curve analysis, as well as with well-established forecasting methods such as ARIMA, neural networks, and some members of the exponential smoothing family. For the study case, a clustering process has been conducted in order to map forecasting difficulty for three clusters of time series. Our experiments evidence that SAM-oil’s is a very competitive method in oil production forecasting: SAM-oil’s forecasts outperform, in average, those from decline curve analysis and the other forecasting methods, both for the clusters and for the whole set of the experimental time series.