Prediksi Energi Panel Surya Menggunakan Metode N-BEATS

Achmad Dani Nursanto, Prof. Dr. Azhari, MT.

2025 | Skripsi | ILMU KOMPUTER

The non-linear, time-dependent nature of solar power complicates its contribution to the grid as intermittency makes day-ahead forecasting all the more necessary, especially with traditional modeling limitations. Therefore, this research implements a day-ahead photovoltaic yield forecasting model via the N-BEATS deep learning architecture, with a tuned XGBoost baseline model for comparison, using real-world data spanning 2017-2023. This work utilizes historical yield predictions and exogenous covariates - irradiance and temperature - as inputs for model tuning, over different time windows, for both models. The test prediction of N-BEATS reports an sMAPE of 8.304 and RMSE of 0.376 on the test set while the baseline XGBoost model reports an sMAPE of 8.771. Thus, the N-BEATS prediction presents an outperforming prediction as N-BEATS is a more effective model for highly temporal, dynamic yields than tree-based models and even serves as a greater-than-average option for reliable solar power management.

The non-linear, time-dependent nature of solar power complicates its contribution to the grid as intermittency makes day-ahead forecasting all the more necessary, especially with traditional modeling limitations. Therefore, this research implements a day-ahead photovoltaic yield forecasting model via the N-BEATS deep learning architecture, with a tuned XGBoost baseline model for comparison, using real-world data spanning 2017-2023. This work utilizes historical yield predictions and exogenous covariates - irradiance and temperature - as inputs for model tuning, over different time windows, for both models. The test prediction of N-BEATS reports an sMAPE of 8.304 and RMSE of 0.376 on the test set while the baseline XGBoost model reports an sMAPE of 8.771. Thus, the N-BEATS prediction presents an outperforming prediction as N-BEATS is a more effective model for highly temporal, dynamic yields than tree-based models and even serves as a greater-than-average option for reliable solar power management.

Kata Kunci : Deep learning, N-Beats, Solar Power, Forecasting