CLASSIFICATION OF PARTIAL SHADING FAILURES IN PV SYSTEMS WITH DISTANCE VARIATIONS USING XGBOOST ALGORITHM

• Main Author: Ayu Wijayanti, Dyah
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/85694
• Institution: Institut Teknologi Bandung

One of the factors that can affect the power production of Solar Power Plants (PLTS) is partial shading, which occurs when part of a PV module is covered by an object above it. There are various types of partial shading that require different handling methods. Therefore, a method to classify the types of partial shading is needed to assist in the maintenance process of PLTS. One of the challenges is the absence of a support structure for the shading object to conduct experiments with partial shading at a distance, meaning the object covering the PV module has a gap from the module. To develop a classification method, research was conducted on the PLTS system at the Center for Advanced Sciences ITB building. In this study, the XGBoost (Extreme Gradient Boosting) machine learning algorithm was chosen, and the Random Forest algorithm was used as a comparison because both use sequential classification tree methods. The comparison results will show which algorithm has the best potential to be used as a classification method in the future. For method development, data was collected using the Solmetric PV Analyzer 1000S tool for normal conditions and six types of shading. The research results yielded a structure design using perforated angle bars that can be used to support the shading object, allowing data collection for shading at a distance. The machine learning classification results also showed that both XGBoost and Random Forest algorithms predominantly used the Vmpp variable with an importance value of >0.3 or 30% for classification. The classification accuracy of XGBoost reached 92% for the test data and 66.8% for the validation data, while Random Forest achieved 96% for the test data and 52% for the validation data, indicating a higher overfitting level in Random Forest. However, both algorithms struggled to distinguish between data with contact shading and data with shading at a distance. This difficulty in distinguishing between contact and distant shading data is the largest factor contributing to the inaccuracy of both algorithms.. Keywords: Solar Power Plants, partial shading, machine learning, XGBoost, Random Forest