Accurate water-cut forecasting is essential in mature water-drive reservoirs, where excessive water production is the primary cause of well shutdowns and reduced field productivity. This study develops an advanced deep-learning framework using Long Short-Term Memory (LSTM) networks to predict water cut in a giant Iranian oil field characterized by strong aquifer support, heterogeneous lithology, and multi-layer completions. A high-resolution dataset comprising 33,697 water-cut measurements was comprehensively preprocessed through duplicate removal, outlier filtering, feature-variance analysis, technical consistency checks, smoothing, and missing-value imputation. After Monte Carlo sensitivity analysis, seventeen static and dynamic features—such as well coordinates, perforation geometry, permeability, oil-rate history, and production time—were selected as model inputs.

A time-consistent data split was applied, where 80% of early-time data were used for training and the remaining 20% for testing to avoid information leakage. The final LSTM architecture, optimized using grid-search hyperparameter tuning, was trained independently for four reservoir zones. The model achieved excellent performance, with Pearson correlation coefficients of 0.957–0.976 on training data and 0.889–0.920 on testing data, while Absolute Relative Error distributions remained tightly centered near zero. The LSTM successfully reproduced long-term production trends, short-term fluctuations, and operational interventions such as recompletions and flow-rate modifications. These results demonstrate that the proposed LSTM-based framework provides a robust and scalable tool for water-cut prediction and supports informed reservoir-management decisions in large heterogeneous water-drive fields