PRODUCTION FORECASTING OF OIL & GAS WELLS IN THE NIGER DELTA USING ARTIFICIAL NEURAL NETWORKS (ANNS)
Faculty
Department
Year of Publication
upload
Publication Type
Abstract
Conventional decline curve analysis (DCA) methods often fail to provide reliable short-term production forecasts in Niger Delta wells due to significant reservoir heterogeneity—permeability variations of up to 300%—as well as frequent operational disruptions, such as pipeline vandalism that can cause annual losses of around 15%. In addition, non‐stationary decline behavior further complicates forecasting efforts. Previous machine learning approaches have not been specifically validated for the unique conditions of the Niger Delta, leaving a gap in
accurate, regionally tailored predictive tools. To address these challenges, we assembled a daily time-series dataset spanning 2010 to 2023 from 10–15 mature wells in Niger Delta fields. Input variables included oil, gas, and water production rates; downhole and tubing pressures; and choke size. We engineered additional features, such as seven‐day rolling means of production rates and time‐since‐last‐workover, to capture temporal dynamics more effectively. Data preprocessing involved robust scaling, linear interpolation to fill missing entries, and removal of outliers beyond three standard deviations. Our predictive model is a feedforward artificial neural network (ANN) with three hidden layers—256, 128, and 64 neurons respectively—using ReLU activations, 20% dropout for regularization, and batch normalization at each hidden layer. We employed a sliding 30‐day input window, optimized the network using the Adam algorithm on mean squared error loss, and implemented early stopping to prevent overfitting. Model performance was evaluated using a chronological split of 70% training, 15% validation, and 15% test data, with metrics including RMSE, MAPE, and R². Results indicate that the ANN consistently outperforms a Random Forest (RF) benchmark across all targets. For oil production, the ANN achieved an R² of 0.92 compared to RF’s 0.87, with an RMSE of 0.34 versus 0.45. The
most pronounced improvement was observed in water rate predictions, where the ANN attained an R² of 0.93 and MAPE of 0.24, while RF achieved 0.83 and 0.35, respectively. Feature‐importance analysis revealed wellhead pressure (correlation r≈0.85 with oil rates) and choke size as key drivers. Furthermore, residual analysis showed that the ANN’s errors maintain uniform variance across the production range, whereas the RF tends to underpredict at higher production
rates. The ANN’s superior performance stems from its ability to capture nonlinear interactions—such as complex choke‐pressure‐fluid relationships—that are critical for modeling sudden surges in water cut. By leveraging engineered features like rolling means and workover timing, we effectively captured temporal dependencies without resorting to recurrent architectures. In practical terms, the high accuracy (R² > 0.90) achieved by the ANN supports proactive choke management and maintenance scheduling, helping operators mitigate downtime in a region notorious for operational interruptions. However, the study has limitations. The dataset is confined to Niger Delta fields, which may limit generalizability across other operators or geological settings. Additionally, the feed forward ANN functions as a “black box,” and further work is required to incorporate explainability methods (e.g., SHAP or layer‐wise relevance propagation) before full field deployment. In conclusion, this research presents the first validated feed forward ANN tailored to Niger Delta wells, enabling reliable 180‐day production forecasts and marking a shift from reactive to predictive production management in this challenging environment.
accurate, regionally tailored predictive tools. To address these challenges, we assembled a daily time-series dataset spanning 2010 to 2023 from 10–15 mature wells in Niger Delta fields. Input variables included oil, gas, and water production rates; downhole and tubing pressures; and choke size. We engineered additional features, such as seven‐day rolling means of production rates and time‐since‐last‐workover, to capture temporal dynamics more effectively. Data preprocessing involved robust scaling, linear interpolation to fill missing entries, and removal of outliers beyond three standard deviations. Our predictive model is a feedforward artificial neural network (ANN) with three hidden layers—256, 128, and 64 neurons respectively—using ReLU activations, 20% dropout for regularization, and batch normalization at each hidden layer. We employed a sliding 30‐day input window, optimized the network using the Adam algorithm on mean squared error loss, and implemented early stopping to prevent overfitting. Model performance was evaluated using a chronological split of 70% training, 15% validation, and 15% test data, with metrics including RMSE, MAPE, and R². Results indicate that the ANN consistently outperforms a Random Forest (RF) benchmark across all targets. For oil production, the ANN achieved an R² of 0.92 compared to RF’s 0.87, with an RMSE of 0.34 versus 0.45. The
most pronounced improvement was observed in water rate predictions, where the ANN attained an R² of 0.93 and MAPE of 0.24, while RF achieved 0.83 and 0.35, respectively. Feature‐importance analysis revealed wellhead pressure (correlation r≈0.85 with oil rates) and choke size as key drivers. Furthermore, residual analysis showed that the ANN’s errors maintain uniform variance across the production range, whereas the RF tends to underpredict at higher production
rates. The ANN’s superior performance stems from its ability to capture nonlinear interactions—such as complex choke‐pressure‐fluid relationships—that are critical for modeling sudden surges in water cut. By leveraging engineered features like rolling means and workover timing, we effectively captured temporal dependencies without resorting to recurrent architectures. In practical terms, the high accuracy (R² > 0.90) achieved by the ANN supports proactive choke management and maintenance scheduling, helping operators mitigate downtime in a region notorious for operational interruptions. However, the study has limitations. The dataset is confined to Niger Delta fields, which may limit generalizability across other operators or geological settings. Additionally, the feed forward ANN functions as a “black box,” and further work is required to incorporate explainability methods (e.g., SHAP or layer‐wise relevance propagation) before full field deployment. In conclusion, this research presents the first validated feed forward ANN tailored to Niger Delta wells, enabling reliable 180‐day production forecasts and marking a shift from reactive to predictive production management in this challenging environment.
Supervisor(s)
co-supervisor