SENSITIVITY ANALYSIS OF HYPERPARAMETER IN SOLAR ENERGY PREDICTION MODEL USING GRADIENT BOOSTING METHOD
| Main Authors: | Ramadhan, Aska, Sopha, Bertha Maya, Ridwan, Mohammad Kholid |
|---|---|
| Other Authors: | Bertha Maya Sopha, Department of Mechanical and Industrial Engineering, Faculty of Engineering, Gadjah Mada University, Mohammad Kholid Ridwan, Department of Nuclear Engineering and Engineering Physics, Faculty of Engineering, Gadjah Mada University |
| Format: | Article info application/pdf eJournal |
| Bahasa: | eng |
| Terbitan: |
Master in Systems Engineering
, 2025
|
| Subjects: | |
| Online Access: |
https://journal.ugm.ac.id/ajse/article/view/78322 https://journal.ugm.ac.id/ajse/article/view/78322/41081 |
Daftar Isi:
- Solar energy prediction is one alternative to handling unpredicted conditions of weather and solar radiation intensity. It could be the most important factor in achieving stability in electricity generation using solar energy resources. In making predictions, the use of machine learning models has been carried out by various methods, and in this study, the method used for the algorithm model is gradient boosting. In the modeling process using gradient boosting, several hyperparameter settings are needed. Hyperparameters have an important role in producing stable predictive patterns and can avoid overfitting or underfitting conditions. In this study, the accuracy and speed of prediction of the machine learning model with the gradient boosting approach, namely XGBoost and LightGBM, were analyzed in relation to setting the hyperparameter learning rate and max depth of the model's prediction pattern. The dataset used spans 6 months at a data resolution rate of every 5 minutes and includes meteorological data at the location point of Energy Laboratory UKRIM Yogyakarta as well as the output value of PLTS power and temperature panels onsite. Setting the hyperparameter learning rate in the highest and lowest conditions generates accuracy values with a difference of 2% and about the same prediction speed. With nMAE values of 2.84% and 1.35% and nRMSE values of 6.11% and 3.68%, respectively, the higher learning rate results in lower error values for both models. The XGBoost model shown tendency for overfitting and slower prediction speeds with the highest max depth setting. The prediction speed is faster at the lowest max depth condition, but the XGBoost and LightGBM models both exhibit underfitting.