Multivariate time-series analysis for long-term forecasting of water cut: a case study for the Zubair AB reservoir unit in the Rumaila oilfield, Southern Iraq

Rising water cut – the fraction of water in produced fluids – poses a significant challenge to sustaining oil recovery in mature reservoirs, yet it has received limited attention in time-series forecasting studies, which have predominantly focused on production rates and price dynamics. In the AB reservoir unit of the Zubair Formation in the Rumaila oilfield, southern Iraq, water cut has become a persistent and increasingly severe production challenge resulting in reduced oil recovery, higher water-handling costs, and more complex reservoir management. Addressing this gap is critical, as reliable long-term water cut forecasts enable operations to optimize injection strategies, sustain production, and avert costly interventions. To this end, the study, investigates the AB unit utilizing a nine-year dataset (2012–2020) of daily oil production, water injection, and water cut records from three producer-injector well groups distributed across the field. A Vector Autoregression (VAR) model was built for each group is jointly forecast water cut five years a head (through 2025), incorporating stationarity testing, Granger-causality analysis, and stability diagnostics; its performance was benchmarked against Autoregressive Integrated Moving Average (ARIMA) models and Long Short-Term Memory (LSTM) deep learning networks. The VAR models demonstrated robust predictive performance, achieving in-sample R 2 values up to 0.99 and out-of-sample test R 2 values between 0.71 and 0.78 across groups. Unlike ARIMA, which underperformed significantly (test R 2 ranging from 0.22 to 0.45), and LSTM, which tended to converge toward a single static prediction, VAR successfully reproduced the non-linear, fluctuating trajectory of water cut across the entire five-year horizon. Furthermore, while LSTM yielded marginally lower error statistics, VAR was uniquely capable of modeling the dynamic, coupled relationships between injection and production that governs water cut behavior. Forecasts diverged by well group: water cut is projected to keep rising in Group I and III, while Group II is projected to decline to roughly 17% by 2025, and Group III to about 30%, reflecting each group’s distinct injection-production dynamics. The principal novelty of this work is the first application of the VAR framework − one that models production, injection, and water cut jointly rather than in isolation − to long-term water cut forecasting, providing reservoir engineers and decision-makers with a practical, interpretable tool for anticipating water cut trends and optimizing injection strategy in the Zubair Formation and comparable heterogeneous reservoirs.