Hybrid XGBoost-LSTM Model for Oil Well Production Forecasting Using the Volve Field Dataset

• A hybrid framework integrates XGBoost for feature extraction with a deep Bidirectional LSTM enhanced by Layer Normalization and Dropout, tailored for short-term oil well production forecasting. • The framework employs engineered lag-based and moving average features to capture temporal patterns and address well-level variability. • Combining gradient boosting for static insights with LSTM for sequential modeling and surpassing standalone models in predictive accuracy. • The framework is computationally efficient, reproducible, and easily adaptable to other oilfields with similarly structured production data. Accurate oil well production forecasting is vital for optimizing reservoir management, mitigating operational risks, and supporting data-driven decision-making in the energy sector. However, daily production data are often noisy and incomplete, and operational interruptions can break temporal continuity, which complicates reliable forecasting. To address these challenges, we propose a novel hybrid forecasting framework that integrates gradient boosting (XGBoost) with a multi-layer bidirectional long short-term memory (Bi-LSTM) network. This two-stage design uses XGBoost as a first-stage nonlinear regressor. Its output is then provided to a BiLSTM using a sliding lookback window so the recurrent layer learns temporal dependencies from multi-day sequences rather than a single-step input. The framework was evaluated on the Volve field dataset using leakage-aware protocols (chronological split and a cross-well holdout test on an unseen well) and was benchmarked against a naïve persistence baseline, XGBoost-only ablation, and standalone LSTM/GRU models. Under the within-well chronological split, the proposed hybrid model achieved MAE = 0.47, RMSE = 3.47, and R 2 = 0.9955 , indicating excellent fit on that specific held-out segment. To provide a more practical and robust assessment, we additionally report unseen-well cross-holdout and walk-forward validation results, together with scale-aware error metrics and an analysis of missing-data sensitivity under forward-fill preprocessing. We report results under two preprocessing settings (i) operational-only samples (ON_STREAM_HRS > 0) and (ii) calendar-time samples including shut-ins to clarify the practical scope of the forecasting task and the impact of operational interruptions. The model’s architecture incorporates dropout and L2 regularization for overfitting control and maintains computational efficiency suitable for real-world digital oilfield deployment. This work validates the synergy between gradient boosting and recurrent neural networks and offers a scalable approach for time series modeling in oil field production analytics.