Peramalan Energi Surya Berdasarkan Dataset Time-Series Menggunakan Model Berbasis Transformer

Al Khwaritsmi, Muhammad Ihsan (2026) Peramalan Energi Surya Berdasarkan Dataset Time-Series Menggunakan Model Berbasis Transformer. Other thesis, Institut Teknologi Sepuluh Nopember.

Text 5025221211-Undergraduate_Thesis.pdf Restricted to Repository staff only Download (30MB) | Request a copy

Abstract

Peningkatan penetrasi energi surya membawa tantangan berupa variabilitas daya akibat kondisi atmosfer dinamis, seperti tutupan awan, yang mengganggu stabilitas jaringan listrik. Pendekatan deep learning seperti LSTM memiliki keterbatasan dalam menangkap dependensi jangka panjang, sehingga arsitektur Transformer mulai diadaptasi untuk peramalan deret waktu. Namun, evaluasi komparatif varian Transformer modern pada peramalan energi surya dengan data multimodal (citra satelit dan meteorologi) masih terbatas. Selain itu, penelitian terdahulu umumnya hanya berfokus pada metrik akurasi simetris dan mengabaikan arah kesalahan prediksi, padahal risiko keamanan operasional berupa overforecasting memiliki dampak dan konsekuensi biaya yang lebih buruk bagi jaringan listrik dibandingkan underforecasting.
Penelitian ini mengevaluasi dan membandingkan kinerja tiga arsitektur berbasis Transformer (PatchTST, iTransformer, dan TimeXer) dengan model baseline CNN-LSTM. Model dilatih menggunakan dataset multimodal dari Toufen Solar Farm, Taiwan (2020), yang mengintegrasikan historis daya, data meteorologi NSRDB, dan data spasial berupa citra inversi satelit beresolusi 10 menit. Metodologi penelitian dievaluasi melalui rolling cross-validation 8 lipatan yang dibagi menjadi empat skenario pengujian berurutan: penentuan fungsi loss terbaik, optimasi lookback window, pencarian hyperparameter menggunakan metode Optuna-TPE, dan penerapan asymmetric loss function (AsymmetricMAELoss) untuk mendorong model melakukan underforecast sebagai mitigasi risiko keamanan operasional. Hasil pengujian pada tahap pertama menunjukkan bahwa fungsi loss MAE secara signifikan (χ² = 50,375, s.(***)) memberikan performa terbaik dibandingkan MSE dan Huber. Pada tahap kedua, ketiga arsitektur Transformer terbukti secara dominan optimal menggunakan lookback window panjang (96 time-steps), berkebalikan dengan arsitektur rekuren LSTM yang kesulitan pada sekuens panjang. Pada tahap tuning, model CNN-PatchTST memperoleh peningkatan performa paling signifikan dengan penurunan MAE hingga 19,23% pada forecast horizon 12. Terakhir, pada pengujian keamanan operasional, CNN-TimeXer terbukti sebagai model yang paling andal, mampu mencapai titik aman dari risiko overforecasting dengan degradasi akurasi yang sangat marginal yaitu sebesar 0,71%, jauh melampaui CNN-LSTM yang membutuhkan penalti besar sehingga terdegradasi hingga 23,83%. Secara keseluruhan, hasil terbaik dan optimal diperoleh oleh model CNN-TimeXer yang menjadi arsitektur paling kompetitif dalam menyeimbangkan akurasi peramalan dan keandalan keamanan operasional.
======================================================================================================================================
The increasing penetration of solar energy introduces power variability challenges due to dynamic atmospheric conditions, such as cloud cover, which disrupt power grid stability. Deep learning approaches like LSTM have limitations in capturing long-term dependencies, prompting the adaptation of Transformer architectures for time-series forecasting. However, comparative evaluations of modern Transformer variants for solar energy forecasting using multimodal data (satellite imagery and meteorology) remain limited. Furthermore, previous studies generally focused on symmetric accuracy metrics and ignored the direction of prediction errors, even though the operational safety risks of overforecasting have more severe impacts and cost consequences for the grid compared to underforecasting.
This study evaluates and compares the performance of three Transformer-based architectures (PatchTST, iTransformer, and TimeXer) against a CNN-LSTM baseline model. The models were trained using a multimodal dataset from the Toufen Solar Farm, Taiwan (2020), which integrates historical power, NSRDB meteorological data, and spatial data in the form of 10-minute inversion satellite imagery. The research methodology was evaluated through an 8-fold rolling cross-validation divided into four sequential testing scenarios: determining the best loss function, optimizing the lookback window, hyperparameter tuning using the Optuna-TPE method, and applying an asymmetric loss function (AsymmetricMAELoss) to encourage the model to underforecast as a mitigation strategy for operational safety risks. The test results in the first stage showed that the MAE loss function significantly (χ² = 50.375, s.(***)) yielded the best performance compared to MSE and Huber. In the second stage, the three Transformer architectures dominantly proved optimal using a long lookback window (96 time-steps), contradicting the recurrent LSTM architecture which struggled with long sequences. During the tuning stage, the CNN-PatchTST model obtained the most significant performance improvement with a MAE reduction of up to 19.23% at forecast horizon 12. Finally, in the operational safety testing, CNN-TimeXer proved to be the most reliable model, achieving a safe point from overforecasting risks with a very marginal accuracy degradation of 0.71%, far outperforming CNN-LSTM which required large penalties resulting in up to a 23.83% degradation. Overall, the best and optimal result was achieved by the CNN-TimeXer model, which is the most competitive architecture in balancing forecasting accuracy and operational safety reliability.

Item Type:	Thesis (Other)
Uncontrolled Keywords:	Deep Learning, Keamanan Operasi, Peramalan Energi Surya, Time-Series, Transformer, Deep Learning, Operational Safety, Solar Energy Forecasting, Time-Series, Transformer.
Subjects:	T Technology > T Technology (General) T Technology > T Technology (General) > T174 Technological forecasting T Technology > T Technology (General) > T57.5 Data Processing
Divisions:	Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Informatics Engineering > 55201-(S1) Undergraduate Thesis
Depositing User:	Muhammad Ihsan Al Khwaritsmi
Date Deposited:	27 Jul 2026 01:32
Last Modified:	27 Jul 2026 01:32
URI:	http://repository.its.ac.id/id/eprint/137722

Actions (login required)

View Item