Selected work

Representative Papers

Forecasting solar power output in Ibadan: A machine learning approach leveraging weather data and system specifications

Aug 10, 2025

Accurate solar irradiance forecasting is critical for solar power planning and grid stability in tropical monsoon regions like Ibadan, Nigeria, yet conventional methods suffer from high dependency on expensive ground-based pyranometers. Method: This study proposes a physics-informed, data-driven two-stage irradiance forecasting framework. Stage one integrates clear-sky models with cloud-type classification to improve hourly global horizontal (GHI), direct normal (DNI), and diffuse horizontal (DHI) irradiance predictions. Stage two employs PVLib to estimate photovoltaic power output using meteorological inputs, cloud conditions, and system-specific parameters. Random forest, CNN, and LSTM models are comparatively evaluated. Contribution/Results: Random forest achieves the best performance, yielding annual normalized root-mean-square errors (nRMSE) of 0.19 (GHI), 0.33 (DNI), and 0.22 (DHI); during the dry season, GHI nRMSE drops to 0.12. The framework significantly reduces reliance on ground measurements and offers a scalable, low-cost solution for high-accuracy solar forecasting in tropical monsoon climates.

0 citationsRead paper

Fusion Sampling Validation in Data Partitioning for Machine Learning

Aug 02, 2025

To address the high computational cost of conventional K-fold cross-validation (KFCV) and the poor representativeness of random sampling on imbalanced datasets, this paper proposes Fusion Sampling Validation (FSV)—a hybrid evaluation framework that adaptively weights Simple Random Sampling (SRS) and KFCV to jointly improve data representativeness and reduce estimation bias. FSV incorporates a scaling factor for bias calibration and monitors convergence rate, enabling superior accuracy-efficiency trade-offs in large-scale and resource-constrained settings. Evaluated via a 5-fold, 10-repetition experimental design, FSV achieves a mean error (ME) of 0.000863, validation accuracy (VE) of 0.9496, MSE of 0.9521, and bias of 0.0163. Results demonstrate that FSV significantly outperforms baseline methods in generalization capability, stability, and convergence speed.

0 citationsRead paper

Recent publications

Latest Papers

Forecasting solar power output in Ibadan: A machine learning approach leveraging weather data and system specifications

Aug 10, 2025

Accurate solar irradiance forecasting is critical for solar power planning and grid stability in tropical monsoon regions like Ibadan, Nigeria, yet conventional methods suffer from high dependency on expensive ground-based pyranometers. Method: This study proposes a physics-informed, data-driven two-stage irradiance forecasting framework. Stage one integrates clear-sky models with cloud-type classification to improve hourly global horizontal (GHI), direct normal (DNI), and diffuse horizontal (DHI) irradiance predictions. Stage two employs PVLib to estimate photovoltaic power output using meteorological inputs, cloud conditions, and system-specific parameters. Random forest, CNN, and LSTM models are comparatively evaluated. Contribution/Results: Random forest achieves the best performance, yielding annual normalized root-mean-square errors (nRMSE) of 0.19 (GHI), 0.33 (DNI), and 0.22 (DHI); during the dry season, GHI nRMSE drops to 0.12. The framework significantly reduces reliance on ground measurements and offers a scalable, low-cost solution for high-accuracy solar forecasting in tropical monsoon climates.

0 citationsRead paper

Fusion Sampling Validation in Data Partitioning for Machine Learning

Aug 02, 2025

To address the high computational cost of conventional K-fold cross-validation (KFCV) and the poor representativeness of random sampling on imbalanced datasets, this paper proposes Fusion Sampling Validation (FSV)—a hybrid evaluation framework that adaptively weights Simple Random Sampling (SRS) and KFCV to jointly improve data representativeness and reduce estimation bias. FSV incorporates a scaling factor for bias calibration and monitors convergence rate, enabling superior accuracy-efficiency trade-offs in large-scale and resource-constrained settings. Evaluated via a 5-fold, 10-repetition experimental design, FSV achieves a mean error (ME) of 0.000863, validation accuracy (VE) of 0.9496, MSE of 0.9521, and bias of 0.0163. Results demonstrate that FSV significantly outperforms baseline methods in generalization capability, stability, and convergence speed.

0 citationsRead paper