| Abstract [eng] |
With the rapid expansion of renewable energy sources, particularly photovoltaics, their integration into national power grids presents significant challenges due to the intermittent nature of solar generation. Accurate forecasting is therefore critical for ensuring grid stability, efficient load balancing, and economic viability. The primary objective of thesis "Daily energy forecasting of solar power plants based on weather forecast data in different locations" is to address the challenge of spatial generalization—predicting solar power generation in new, previously unmonitored locations where historical generation data is unavailable, and to determine the impact of various meteorological variables on prediction accuracy. To achieve this, the study investigates and compares two distinct machine learning approaches: the deep learning-based Long Short-Term Memory (LSTM) neural networks, designed to capture temporal dependencies, and the ensemble-based eXtreme Gradient Boosting (XGBoost) decision trees. The research utilizes a dataset comprising historical generation data from 11 distinct solar power plants alongside open-access meteorological reanalysis data from "Open-Meteo". To enhance model performance, extensive feature engineering was performed, including the calculation of panel variables (azimuth and tilt angles). Furthermore, the Hyperband optimization algorithm was used to tune hyperparameters and select the optimal model architectures. The experimental phase involved training models on a dataset from 10 power plants and testing their performance on an 11th, "unseen" plant to simulate a real-world scenario of expanding to a new site. The results demonstrated that the XGBoost model significantly outperformed the LSTM network, achieving approximately 15 % lower MAE on average while requiring substantially lower computational resources for training. The study highlights a trade-off between generalization and precision: comparative analysis showed that models trained specifically on local data from a single plant still yield higher accuracy than generalized models. The thesis concludes with recommendations for further research, suggesting that integrating real-time satellite imagery or ground-based cloud observation data could further mitigate errors caused by cloud movement. |
