Proposal For A Performance Evaluation Method For Grid-Connected Photovoltaic Systems Based On Machine Learning Forecasting Models

In this work, the authors present the results of a study conducted at the University of Bras´ılia (UnB), Brazil, which had as its main objective to evaluate and compare daily forecast models for electricity generation in photovoltaic (PV) plants. The proposed solution is based on an infrastructure for mea- suring electrical quantities and developing supervisory software, enabling continuous operational monitoring of the University’s buildings. The model evaluation employs six standardized metrics for solar energy forecasting, following NREL guidelines and best international practices. These metrics enable objective selection of the most suitable algorithms for operational photovoltaic plants applications, providing a comprehensive assessment of both accuracy and reliability. The top three exogenous models provided nMAE (Normalized Mean Absolute Error) around 4.91-5.10%, which significantly outperform the three best tem- poral models, which provided nMAE around 14.16-14.27%, with an improvement of about 65% in forecast error. Even more critical, the coefficient of determination (R2) rose from 0.12 to 0.90, demonstrating that physical variables encode the essential mechanisms of PV energy generation that temporal variables cannot adequately represent. These models represent the three most effective features of engineering strategies. Only models that presented nMAE below 5.2% and R2 above 0.89 were chosen, ensuring high forecasting accuracy for reliable operational anomaly detection. An anomaly detector based on the Energy Performance Index (EPI), framework defined in IEC 61724, was developed using a machine learning model and used as the expected performance reference, which normalizes the operational analysis by removing the effect of weather conditions. The exceptional accuracy of the selected models (nMAE 5%, R² > 0.89) shows that the proposed solution has great potential for application in Energy Management Systems (EMS), as it enables the timely implementation of preventive and corrective maintenance measures, avoiding financial losses.