MACHINE LEARNING ALGORITHMS

This project presents the optimization of solar inverter efficiency using machine learning algorithms to improve power generation accuracy and system reliability under varying environmental conditions. Traditional solar inverter systems and Maximum Power Point Tracking (MPPT) methods often experience limitations in adapting to fluctuations in solar irradiance, temperature, and shading conditions, leading to reduced efficiency and energy loss. To address these challenges, this study developed and evaluated machine learning models capable of predicting and optimizing inverter performance in real time. Environmental and operational data including irradiance, temperature, day, hour, and inverter performance metrics were collected from the NASA and NSRDB datasets for the University of Benin region. Data preprocessing techniques such as normalization, interpolation, and feature engineering were applied before model training. Three machine learning models — Random Forest (RF), Gradient Boosting Machine (GBM), and Artificial Neural Network (ANN) — were implemented and evaluated using Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), and Coefficient of Determination (R²). Results showed that the ANN model outperformed the other models with an MAE of 0.019, RMSE of 0.029, and R² value of 0.962. The optimized system achieved an efficiency improvement of 8.3% compared to conventional MPPT methods. The study further demonstrated the capability of machine learning algorithms to adapt to changing environmental conditions and improve solar inverter performance. The developed model was deployed using Django REST Framework for real-time prediction and monitoring. This research confirms that machine learning-based optimization can significantly enhance solar inverter efficiency, reduce energy losses, and contribute to sustainable and intelligent renewable energy systems.

The accurate prediction of the Well roductivity Index (PI) is critical for reservoir management, production optimization, and forecasting. Traditional methods, such as analytical correlations and decline curve analysis, are often limited by simplifying assumptions that fail to capture the complexities of heterogeneous reservoirs like those in the Niger Delta. This research addresses
this limitation by developing and evaluating a data-driven framework for PI prediction using machine learning (ML) on historical production data. The study implements and compares three advanced ensemble regression algorithms—Random Forest, GBoost, and CatBoost—to predict PI from daily records of oil, gas, and water production rates and downhole pressures. A dataset approximately 7,000 daily records from five Niger Delta wells was utilized, with the PI target variable calculated using a proxy for reservoir pressure drawdown. A clear performance hierarchy was established among the models. Random Forest yielded the weakest performance (R² = 0.18, MAE = 65.62), while XGBoost showed substantial improvement (R² = 0.78, MAE = 34.14). CatBoost erged as the superior model, achieving exceptional predictive accuracy with an R² of 0.95, a Mean Absolute Error (MAE) of 18.96, and a Root Mean Squared Error (RMSE) of 21.02. Residual and temporal analyses confirmed that CatBoost produced unbiased, homoscedastic errors and effectively tracked the dynamic PI trends of individual wells over time. Interpretability analyses revealed that production rates (oil, gas, and water) were the most influential predictors, a finding consistent with reservoir engineering principles. However, this also highlights a methodological caveat regarding the mathematical coupling between the model's inputs and the PI target. The study concludes that CatBoost provides a robust and highly accurate model for PI prediction from routine field data, offering a significant advantage over traditional methods for well performance monitoring and screening in the Niger Delta context

The accurate prediction of the Well Productivity Index (PI) is critical for reservoir management, production optimization, and forecasting. Traditional methods, such as analytical correlations and decline curve analysis, are often limited by simplifying assumptions that fail to capture the complexities of heterogeneous reservoirs like those in the Niger Delta. This research addresses this limitation by developing and evaluating a data-driven framework for PI prediction using machine learning (ML) on historical production data. The study implements and compares three advanced ensemble regression algorithms—Random Forest, XGBoost, and CatBoost—to predict PI from daily records of oil, gas, and water production rates and downhole pressures. A dataset of approximately 7,000 daily records from five Niger Delta wells was utilized, with the PI target variable calculated using a proxy for reservoir pressure drawdown. A clear performance hierarchy was established among the models. Random Forest yielded the weakest performance (R² = 0.18, MAE = 65.62), while XGBoost showed substantial improvement (R² = 0.78, MAE = 34.14). CatBoost emerged as the superior model, achieving exceptional predictive accuracy with an R² of 0.95, a Mean Absolute Error (MAE) of 18.96, and a Root Mean Squared Error (RMSE) of 21.02. Residual and temporal analyses confirmed that CatBoost produced unbiased, homoscedastic errors and effectively tracked the dynamic PI trends of individual wells over time. Interpretability analyses revealed that production rates (oil, gas, and water) were the most influential predictors, a finding consistent with reservoir engineering principles. However, this also highlights a methodological caveat regarding the mathematical coupling between the model's inputs and the PI target. The study concludes that CatBoost provides a robust and highly accurate model for PI prediction from routine field data, offering a significant advantage over
traditional methods for well performance monitoring and screening in the Niger Delta context.