AI Ensemble Outsmarts Classic Decline Curves in Niger Delta Oil Forecasting

Forecasting how much oil a mature field will produce has always been part science, part educated guesswork. Now a team at the University of Port Harcourt in Nigeria has shown that a carefully orchestrated combination of machine learning models, classical reservoir engineering, and a large language model can outperform the traditional tools of the trade. In a study published in Discover Geoscience, Fossong Guilianno and colleagues analysed thirty-two years of production data from the Gabo Field in the Niger Delta and generated five-year forecasts that were consistently more reliable than those from conventional decline curve analysis alone.

The Niger Delta is one of the world’s most prolific petroleum provinces, but its reservoirs are notoriously difficult to model. Interbedded sands and shales create compartmentalised flow units, and production behaviour can shift abruptly as water breaks through into high-permeability channels or as pressure depletes unevenly across the field. The classic engineering answer to this problem, decline curve analysis, fits mathematical decline equations to historical production rates and extrapolates them forward. That approach assumes a smooth, predictable decline, which is precisely what heterogeneous reservoirs refuse to deliver. The new study set out to quantify just how much those assumptions cost in accuracy, and to build a framework that could recover the lost information.

The researchers assembled monthly records of oil, gas, and water production rates, along with gas-oil ratios, from six wells covering 1992 through 2024. Before any modelling began, the data passed through a rigorous preprocessing pipeline. Missing timestamps were filled by linear interpolation, physically impossible negative production values were clipped to zero, and the series were log-transformed to stabilise their variance. The team then engineered temporally causal features, including lagged production values, rolling statistical summaries, and cyclical seasonal encodings, all constructed so that no future information could leak into the training process. Wells with fewer than six valid monthly observations were excluded entirely, because such short sequences cannot support reliable temporal learning.

Three fundamentally different forecasting engines formed the core of the system. A Random Forest regressor learned nonlinear relationships from the engineered lag features, using between 200 and 300 decision trees. A Long Short-Term Memory neural network, with two stacked recurrent layers of 64 and 32 hidden units, processed twelve-month sequences of raw production history to capture annual seasonality and medium-term dependencies. Facebook Prophet, a Bayesian structural time-series model, decomposed the data into trend and seasonal components. Each model was trained and evaluated under identical expanding-window validation, in which the most recent twelve months of each well were held out as a test set, mimicking the real-world condition in which forecasters never see the future.

The crucial innovation came in how the three models were combined. Rather than simply averaging their predictions, the team used XGBoost stacking, a meta-learning technique in which the individual model outputs for each validation timestamp were fed as features into a gradient-boosted regressor that learned the optimal way to blend them. Because the meta-learner was trained only on out-of-fold predictions generated from data unseen during base-model training, the stacking avoided temporal misalignment and target leakage, two of the most common failure modes in applied time-series work. A simpler weighted-averaging ensemble, with weights inversely proportional to each model’s validation error, served as a comparison point.

The results were striking. The stacked ensemble achieved coefficients of determination between roughly 0.97 and 0.98 in the transformed forecasting space, and its Mean Absolute Scaled Error fell below 1 for most wells, meaning it beat the naive persistence benchmark that simply predicts tomorrow’s production equals today’s. The Random Forest on its own reached an R-squared of about 0.89, matching benchmarks reported in the wider literature, while the ensemble’s validation errors in original field units were lower than any single model across the six wells. Hyperbolic decline models, the most flexible of the classical equations, still fit the data better than exponential ones, with a root mean square error of 22.4 barrels per day versus 34.7, confirming that Gabo’s wells sit in transitional rather than boundary-dominated flow regimes.

Not every model performed equally well, and the failures were as informative as the successes. The LSTM network degraded on wells with short production histories, a reminder that deep recurrent networks need long, stable sequences to learn from. Prophet stumbled badly on one well, Gabo-35, whose abrupt fluctuations and irregular declines violated the additive trend-and-seasonality assumptions the model is built on. The authors also drew an important distinction between metrics computed in the log-transformed space, where errors appear tiny, and metrics in original production units, where inverse transformations amplify deviations at high flow rates. Only the original-scale numbers, they argue, are operationally meaningful for reservoir management decisions.

The study’s most unusual element was its use of DeepSeek-R1, a large reasoning model, as a cognitive interpretation layer rather than a forecasting engine. Structured outputs from the machine learning pipeline, along with decline-curve parameters, water-cut trends, and gas-oil ratio statistics, were fed to the model running locally, which then flagged recurring operational patterns. Its analysis revealed that 23 percent of the wells showed irregular production declines exceeding 50 percent on a monthly basis, and it identified three recurrent signatures: early-life instability from pressure depletion, water breakthrough along high-permeability channels, and late-life variability from compartmentalisation. These interpretations were cross-checked against observed production behaviour and known geology of the Gabo Field, and the authors are careful to describe the cognitive layer as qualitative decision support, not validated prediction.

The practical implications reach well beyond one field. The five-year cumulative production projections from the stacked ensemble ranged from about 539,000 stock tank barrels at Gabo-42 to roughly 4.6 million at Gabo-18, giving operators concrete numbers for development planning. The team estimates that conventional Arps-style decline analysis can undervalue late-life production by up to 18 percent, a gap large enough to distort reserve estimates and drilling economics. Wells flagged with irregular declines, they suggest, could be prioritised for selective recompletion or targeted pressure support, and the hybrid framework reportedly cut twelve-month forecast errors by 18 percent compared with existing CNN-LSTM implementations.

Limitations remain, and the authors are candid about them. Fifteen percent of the gas-oil ratio measurements exceeded thermodynamic solubility limits, signalling that pressure-volume-temperature data need recalibration before they can be fully trusted. Forecast uncertainty inevitably grows as the horizon extends beyond observed data, and the reliability of AI-generated reservoir interpretations has yet to be formally benchmarked against expert judgment. The researchers propose several next steps: physics-informed models that couple material balance equations with neural forecasts, online learning systems that update parameters as new production data arrive, and the integration of seismic and geological information to add spatial resolution. If those directions mature, the combination of ensemble machine learning and cognitive computing could move the industry a significant step closer to autonomous reservoir management, in which fields effectively monitor, interpret, and optimise themselves.

Subject of Research: Ensemble machine learning and cognitive computing for oil production forecasting in a heterogeneous Niger Delta reservoir

Article Title: Oil production forecasting in a heterogeneous Niger Delta reservoir using ensemble machine learning and cognitive computing

Article References: Guilianno, F., Okengwu, K. O., & Okengwu, U. A. (2026). Oil production forecasting in a heterogeneous Niger Delta reservoir using ensemble machine learning and cognitive computing. Discover Geoscience, 4(1), Article 317. https://doi.org/10.1007/s44288-026-00688-y

Image Credits: AI Generated

DOI: 10.1007/s44288-026-00688-y

Keywords: oil production forecasting, Niger Delta, ensemble machine learning, XGBoost stacking, LSTM, Random Forest, Prophet, decline curve analysis, DeepSeek-R1, reservoir heterogeneity, Gabo Field, cognitive computing