Short-term solar forecasting in Germany

: using satellite imagery and a hybrid CNN-2-LSTM approach

Abstract

This thesis presents a novel hybrid convolutional neural network to long short-term memory (CNN-2-LSTM) model for short-term solar forecasting using data from Southeast Germany. The model’s performance is compared against a long short-term memory (LSTM) model and a persistence forecast. Satellite-derived cloud masks from the CLAAS-3 dataset by METEOSAT are combined with minute-byminute global horizontal irradiance (GHI) data from the Fraunhofer Institute’s PV-Live dataset. The models predict solar output over 15 minutes, 3 hours, and 6 hours with various input sequence lengths, resulting in 37 models tested. A new performance metric is introduced, using balance energy prices to calculate the economic impact of the models, providing a practical perspective on financial implications. Results show that the CNN-2-LSTM model significantly outperforms benchmarks at the 15-minute horizon, demonstrating superior accuracy for very short-term predictions. However, its performance at the 3-hour horizon is comparable to the LSTM model, and it lags behind the LSTM model at the 6-hour horizon. These findings highlight the model’s effectiveness for very short-term forecasting while emphasizing the need for optimization for specific temporal scales. The research underscores the potential of hybrid deep learning approaches to enhance short-term solar forecasting accuracy. It offers a cost-effective alternative to more complex systems, valuable for solar energy businesses. The study encourages further exploration into optimizing deep learning models for different forecasting horizons, contributing to advancements in renewable energy management in line with sustainable development goals (SDG).

Date of Award	4 Jul 2024
Original language	English
Awarding Institution	Universidade Católica Portuguesa
Supervisor	Pedro Afonso Fernandes (Supervisor)