Enhancing the Performance of Mabushi Solar Power Plant through Machine Learning Forecasting Models: Generation and Demand Sides Management

Abstract

This study explores enhancing the performance of Mabushi Solar Power plant through machine learning forecasting models: Generation and Demand Side Management. The research aims to develop models that can accurately forecast daily solar power output using environmental and operational data. A dataset of 108 daily records within 2023 to part of 2024 from the Mabushi solar power plant was used, incorporating variables such as solar power output (as dependent variable), irradiance, temperature, wind speed, humidity as (independent variables), and energy contributions from batteries, generators, and the grid. The methods adopted are data collection and processing, model development and training. Three machine learning models Random Forest, Gradient Boosting, and Prophet were developed and evaluated. Among these, the Gradient Boosting model proved most effective in optimal forecasting, achieving a Mean Absolute Error (MAE) of 0.006 MW, Root Mean Square Error (RMSE) of 0.008 MW, and a high R² score of 0.999. Its residual errors were tightly distributed around zero, indicating strong reliability and accuracy. The model was then used to forecast solar power generation into the future, with average daily predicted power output of 0.677 MW, with values ranging between 0.390 MW and 1.002 MW, culminating in a total predicted annual generation of approximately 247.04 MW. By applying ML algorithms to 365 days of data, this research produced accurate and interpretable forecasts of daily solar power generation. The results provided actionable insights for load scheduling, maintenance planning, and energy storage management. The forecast makes the plant predictive and not reactive, showcasing its practical relevance for energy planning and operational decision-making through demand side management.