Institutional Repository of Vilnius University
Saulės fotovoltinių elektrinių valandinės energijos prognozavimo modelių ir sezoninių dėsningumų analizė /
Abstract
dc:descriptionThis study investigates the challenge of solar energy production forecasting by applying machine learning and deep learning models. The aim of the research is to evaluate the accuracy of hourly forecasting models for solar photovoltaic (PV) plants, identify the most significant features used for forecasting, and analyze the influence of seasonal patterns on the results. Three datasets were used, gotten from three grid-connected solar PV plants located in different regions of Lithuania. Data preprocessing involved removing rows with missing values and ensuring a chronological sequence to prevent temporal data leakage. Additionally, days with fewer than 12 hours of observations were excluded. Following feature engineering, the final datasets contained 104 features. Five models were selected for day-ahead forecasting: support vector machine (SVM), random forest (RF), extreme gradient boosting (XGBoost), long short-term memory (LSTM) neural network, and the temporal fusion transformer (TFT). Hyperparameter optimization was performed for each model, using expanding window cross-validation for the machine learning models and the LSTM neural network, while the Optuna library was applied to the TFT model. Model accuracy was evaluated using RMSE, MAE, HWMAPE, and DWMAPE metrics. The results indicated that when using the full feature set, the TFT model achieved the lowest average HWMAPE error of 27.24%, while the XGBoost model achieved the lowest DWMAPE error of 10.93%. To reduce data dimensionality and assess feature significance, the Borda rank aggregation method was applied, enabling the selection of the top 20 most significant features. Solar radiation, temperature, and their interaction terms were identified as the most significant features. Temperature variations, air humidity, rolling averages of radiation, and relative indicators also proved to be significant. Feature selection improved the accuracy of 57 % of all experiments, with the most substantial positive impact observed in neural network models. Using the selected features, the TFT model achieved the best and most stable HWMAPE result (μ = 23.42% and σ = 2.97 %), while XGBoost was considered the best model for DWMAPE due to its stability (σ = 1.63%) and low average error (μ = 11.73%). The seasonal analysis revealed that the largest forecasting errors occurred during winter due to minimal energy production and a lack of informative data variance, whereas the best results were achieved in summer, when the data exhibited a regular periodicity. Furthermore, the experiment with the combined global dataset demonstrated that the XGBoost and TFT models can effectively capture shared patterns across different sites. This eliminates the need to develop and maintain separate, individual models for each power plant while maintaining the same performance trend: the TFT model proved to be the most suitable for hourly solar energy forecasting (global HWMAPE = 23.53 %), while the XGBoost model was optimal for daily energy amount prediction (global DWMAPE = 11.52 %).
Degree
thesis:*- Grantor dc:publisher
- Institutional Repository of Vilnius University
- Year dc:date
- 2026
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Semionov, Ratmir,
Rights
dc:rights- Statement dc:rights
-
- info:eu-repo/semantics/openAccess
- Language dc:language
- lit
Identifiers
dc:identifier.*- Repository record dc:identifier
- https://repository.vu.lt/VU:ELABAETD308334984&prefLang=en_US