Probabilistic Time Series Forecasting for Unconventional Oil and Gas Producing Wells

Abstract

Time-series forecasting, the process of predicting values in the future given the present and previous history, is a challenging problem to tackle. Deterministic forecasting methods were thoroughly investigated but had limitations regarding reliability. Recent research efforts are exploring the advantages that come with probabilistic forecasting. The need to have large datasets for time-series to build more generalized models and thus being less dependent on data augmentation is also driving efforts to collect comprehensive data. This paper proposes a machine learning model to estimate prediction intervals on a large oil production dataset. Prediction intervals are estimated at different percentiles. Prediction Interval Coverage Probability (PICP) and Prediction Interval Normalized Average Width (PINAW) metrics are used for performance evaluation. The best results are obtained by removing trend and using differencing.