IMPROVEMENT OF THE POWER PREDICTION MODEL FOR SOLAR POWER PLANT ON MICROGRID SYSTEM USING DEEP NEURAL NETWORK MODEL AND EVENT-TRIGGERED UPDATE METHOD

• Main Author: Waldri, Thoriq
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/85690
• Institution: Institut Teknologi Bandung

Currently regulations around the world emphasize the use of renewable energy (RE) and reducing the use of fossil energy. RE that is being widely used at this time is solar energy, especially in Indonesia, because the potential for solar energy in Indonesia is quite high. However, the fluctuating characteristics of solar energy cause problems in the energy management process. For this reason, the microgrid system comes as a solution, by enabling the coupling of solar energy with the conventional power grid to form a more efficient and reliable energy distribution system. With the utilization of machine learning (ML) technology, can help creating a digital model of a microgrid system to predict the power generated by the Solar Power Plant, such as the one that has been implemented in Energy Management Laboratory, Engineering Physics Study Program, Bandung Institute of Technology. In this final project research, improvement of the model for predicting the output power of solar power plant using Deep Neural Network (DNN) learning model learning with Event-Triggered Update (ETU) method. Improvement action is carried out on the basis that the current prediction model is less relevant and has a very large prediction error. Improvement of the model in the form of modification of hyperparameters, addition of Angle of Incidence (AoI), addition of operating modes and the addition of Custom Loss Function (CLF). Three variations of the improvement model have been made, namely models A, B and C. Based on the results of the improvement model evaluation on the test data during June 2024, model C was the best model with a significant decrease in nRMSE and nMAE of by 27.84% and 34.10% when compared to the current model. Model C has the highest R2 value and produces a relatively small number of negative predictions. In addition, the energy production error during model testing is very small, at 2.34%. The results of evaluating model C with the ETU method obtained quite good results, where model C can adapt if there are nMAE and nRMSE values that exceed the predetermined tolerance limits. predetermined tolerance limits. Keywords: renewable energy, microgrid, improvement, solar power plants, evaluation, deep neural network, event-triggered update. ?