Peramalan Daya Listrik Tenaga Surya Berdasarkan Data Fusi Citra Satelit Dan Data Meteorologi Menggunakan Model CNN-LSTM

Mahendra, Putu Indra (2026) Peramalan Daya Listrik Tenaga Surya Berdasarkan Data Fusi Citra Satelit Dan Data Meteorologi Menggunakan Model CNN-LSTM. Other thesis, Institut Teknologi Sepuluh Nopember.

Text 5025221215_Undergraduate_Thesis.pdf Restricted to Repository staff only Download (4MB) | Request a copy

Abstract

Transisi energi surya menghadapi tantangan stabilitas pasokan akibat fluktuasi cuaca yang dinamis, khususnya di wilayah dengan topografi kompleks seperti Miaoli, Taiwan. Untuk mengatasi masalah tersebut, pemodelan fusi data multimodal yang menggabungkan citra satelit dan parameter cuaca numerik sering diusulkan sebagai sistem peramalan jangka pendek. Namun, integrasi fitur visual ke dalam model belum tentu menjamin peningkatan akurasi di berbagai kondisi cuaca. Oleh karena itu, penelitian ini bertujuan untuk membandingkan secara objektif tingkat akurasi prediksi antara arsitektur fusi data (citra dan tabular) dengan pemodelan deret waktu tabular murni. Penelitian ini membangun sistem peramalan berarsitektur hibrida CNN-LSTM dengan skema Multiple-Input Multiple-Output (MIMO) untuk memprediksi keluaran daya listrik tiga langkah ke depan secara simultan. Setelah prapemrosesan data citra satelit Himawari-8 dan data tabular historis, optimasi parameter dilakukan menggunakan algoritma Tree-structured Parzen Estimator (TPE). Evaluasi model dieksekusi melalui tiga tahapan: seleksi pengekstraksi fitur citra terbaik, optimasi penyesuaian dimensi fusi (Learned Feature Fusion), dan pengujian akhir yang membandingkan model fusi multimodal yang diajukan dengan model LSTM tabular murni. Hasil pengujian menunjukkan bahwa penggabungan data citra dan tabular secara mentah merusak akurasi dengan lonjakan RMSE hingga 57,92 kWh. Meskipun metode Learned Feature Fusion berhasil menyeimbangkan porsi informasi dan menekan RMSE menjadi 40,62 kWh, model LSTM tabular murni menunjukkan performa yang lebih baik pada komparasi akhir dengan mencatatkan rata-rata RMSE sebesar 28,15 kWh. Integrasi fitur spasial citra satelit dengan metode learned feature fusion pada studi kasus ini cenderung memicu redundansi dan lebih rentan terhadap data citra yang terputus. Dengan demikian, model LSTM tabular memiliki rata-rata RMSE yang lebih stabil dan unggul di hampir setiap musim dibandingkan dengan arsitektur fusi multimodal yang diterapkan dalam peramalan jangka pendek.
===================================================================================================================================
The solar energy transition faces supply stability challenges due to dynamic weather fluctuations, especially in regions with complex topography such as Miaoli, Taiwan. To overcome this problem, multimodal data fusion modeling that combines satellite imagery and numerical weather parameters is often proposed as a short-term forecasting system. However, the integration of visual features into the model does not necessarily guarantee an improvement in accuracy under various weather conditions. Therefore, this study aims to objectively compare the prediction accuracy level between data fusion architecture (image and tabular) and pure tabular time-series modeling. This study builds a forecasting system with a hybrid CNN-LSTM architecture using a Multiple-Input Multiple-Output (MIMO) scheme to predict electrical power output three steps ahead simultaneously. After preprocessing the Himawari-8 satellite imagery data and historical tabular data, parameter optimization is performed using the Tree-structured Parzen Estimator (TPE) algorithm. The model evaluation is executed through three stages: selection of the best image feature extractor, optimization of fusion dimension adjustment (Learned Feature Fusion), and final testing that compares the proposed fusion model with the pure tabular LSTM model. The test results show that the raw combination of image and tabular data degrades accuracy with an RMSE spike up to 57.92 kWh. Although the Learned Feature Fusion method successfully balances the information proportion and suppresses the RMSE to 40.62 kWh, the pure tabular LSTM model shows better performance in the final comparison by recording an average RMSE of 28.15 kWh. The integration of satellite image spatial features using the learned feature fusion method in this case study tends to trigger redundancy and is more vulnerable to missing image data. Thus, the tabular LSTM model has a more stable and superior average RMSE in almost every season compared to the applied fusion architecture.

Item Type:	Thesis (Other)
Uncontrolled Keywords:	CNN, Deep Learning, Feature Fusion, LSTM, Peramalan Energi Surya. ======================================================================================================================== CNN, Deep Learning, Feature Fusion, LSTM, Solar Power Forecasting.
Subjects:	T Technology > T Technology (General) > T174 Technological forecasting
Divisions:	Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Informatics Engineering > 55201-(S1) Undergraduate Thesis
Depositing User:	Putu Indra Mahendra
Date Deposited:	28 Jul 2026 02:14
Last Modified:	28 Jul 2026 02:14
URI:	http://repository.its.ac.id/id/eprint/138058

Actions (login required)

View Item