Case Study: Photovoltaic Power Forecasting in Cold Regions and Potential Improvements Through Machine Learning

While solar energy holds great promise in addressing climate change and reducing reliance on fossil fuels, its adoption is hindered by inherent variability due to weather conditions and environmental factors. As a potential solution to address these challenges, we developed a multi-source data integration framework utilizing real power measurements from an installed photovoltaic (PV) system, simulation outputs, and a machine learning (ML) model configured to predict and correct the error between simulated and measured power outputs. The ML algorithm leverages a stacking ensemble method to effectively account for complex environmental interactions, such as snow accumulation, temperature, and wind effects leading to significant improvements in forecasting reliability under diverse conditions. By uncovering complex relationships between environmental factors, our approach allows for adjustments during days and months with snowfall, addressing scenarios where traditional forecasters may overlook local phenomena such as snow coverage. Through this project, we utilized ML to predict the discrepancy (error) between simulated and actual power generated by the PV system. Using the developed model in conjunction with a power forecaster provided data that demonstrated the model's ability to minimize prediction errors compared to real power for the PV system installed at South Dakota Mines.