Ultra-Short-Term Solar PV Power Forecasting Using Gradient Boosting Trees
Pysyvä osoite
Kuvaus
Opinnäytetyö kokotekstinä PDF-muodossa.
As solar PV power generation becomes more significant in the Nordic power systems, opera-tion, flexibility management and local energy management are also becoming more challeng-ing. In high-latitude regions, such as Helsinki, Finland, cloud-induced irradiance fluctuations, large sea-sonal variations in Solar elevation, shorter daylight in winter and snowy/icy surface conditions can greatly influence PV production, making power forecasting more challenging. These conditions lead to forecast uncertainty at very short time horizons and demand fore-casting techniques with the ability to forecast both fast power variations and seasonal operat-ing patterns. The aim of this thesis is to examine the possibilities of ultra short-term (UST) PV power forecasting with a 10–60-minute forecast horizon for the Helsinki Kumpula measure-ment site data. A supervised machine-learning approach using Gradient Boosting Regression Trees (GBRT) is designed to model the non-linear relationship between PV power output and environmental predictors. The input features are lagged PV power measurements, irradiance-related features, cloud cover, air temperature, wind speed, relative humidity, solar-geometry features, clear-sky normalization, and snow-sensitive predictors. These features are intended to depict the most recent system behavior, evolving me-teorological conditions, and seasonal effects in the high latitudes. A time-series-aware validation approach is used to evaluate model performance, and a persistence benchmark is used for com-parison. RMSE, NRMSE, MAE, R2 and forecast skills are used to measure forecast accuracy. The outcomes show that the model results in the best accuracy for the smallest forecast time used in the experiment, 10 minutes, with an RMSE of 1.94 kW and an nRMSE of 10.03%. The RMSE and nRMSE values are also slowly increasing as the prediction horizon increases, increasing to 2.41kW and 12.46% re-spectively after 60 minutes. The GBRT model is superior to the persistence in fore-casting skills as seen from the skill scores of 0.22 at 10 minutes and 0.27 at 60 minutes, which are positive values. The results also indicate that recent PV power observations are more significant than the other predictors for short forecast leads, while the meteorological, solar-geometry, and snow-related predictors gain in significance with increasing forecast leads. The central idea of this thesis is a systematic and transparent forecasting framework based on GBRT for the ultra-short-term prediction of PV power in Nordic high latitude conditions. The proposed solution enhances the short-term situational awareness of PV generation, thus paving the way for the seamless inte-gration of solar power into flexible power systems. The results also underpin future hybrid ap-proaches to forecasting, based on machine learning and physical irradiance modelling and sky-camera-based cloud information.