SOLAR POWER SYSTEM PRODUCTION FORECASTING USING INFORMED MACHINE LEARNING METHOD

• Main Author: Nur Muhammad, Fikri
• Format: Theses
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/70628
• Institution: Institut Teknologi Bandung

Increasing the production of electrical energy produced by solar power through Photovoltaic (PV) presents new challenges in maintaining the stability of the electricity network due to its variability and intermittent nature which depends on weather conditions. Prediction of electrical energy production is a very important energy management feature for a microgrid energy management system so that it can ensure an uninterrupted supply of electrical power to loads in the future. Several studies have attempted to utilize Machine Learning to predict the Solar Power System (PLTS), this is because Machine Learning, especially in the use of Deep Learning, is increasingly important and has been successful in various fields of engineering and science. Despite its tremendous success in making predictions, Machine Learning has limitations when it comes to dealing with insufficient training data. A potential solution would be additional integration of prior knowledge incorporated into the training process and lead to ideas for Informed Machine Learning methods. Therefore, it is proposed to predict electrical energy from the PLTS system using the Informed Machine Learning method in this study, by adding input weather data and PLTS energy production data and carried out in several test scenarios with each test scenario divided into 2 (two) different timestamps, every minute and every hour. This research is one of the digital model deployments to support the Digital Twin performance framework. In carrying out prediction models, the Big Data Analytics infrastructure is a reference in this research. Measuring the performance value of the prediction model uses MAE (Mean Absolute Error), RMSE (Root Mean Square Error), and NRMSE (Normalized Root Mean Square Error) values. Using the Informed Machine Learning method and tuning the Deep Neural Network (DNN) model, the best results are MAE 0.23 and RMSE 0.34 (for timestamp set per minute) and MAE values 0.23 and RMSE 0.29 (for timestamp set per hour). In addition, it was also found that the average error reduction from reference research for MAE was 80.33% and RMSE was 79.42%. The results of the prediction model with the average NRMSE performance value show that the NRMSE with an hourly timestamp is 0.67%. this number is better than minutely timestamp with 0.96% result.