A Time-Driven Approach Leveraging Universally Accessible Features for PV Power Forecasting

Accurate hourly photovoltaic (PV) power forecasting is important for integrating renewable energy into power grids. However, it is an operation inherently limited by the uncertainty of meteorological data. Previous operational performance studies also show installations with considerable inter-annual deviations. This prompted the need for a more robust forecasting methodology based on highly reliable and accessible time-based features, such as hour of day and day of year. These can substitute meteorological inputs if they show consistent associations with PV generation outputs. To develop this method, we used hourly PV generation data from an industrial site in Konstanz, Germany, openly shared by the Open Power System Data project. We created a dataset that combined time-based features with the original meteorological features. For input selection, we prioritized features according to availability and reliability through correlation analysis and assessments of interdependencies. We used diverse forecasting models in our analyses. The LSTM + LGBR model served as our target hybrid model. The forecasting model based on time-based inputs produced R²=۰.۹۷۶۵, MAE=۰.۰۰۵۳, MSE=۰.۰۰۰۳, and RMSE=۰.۰۱۵۹. These results were closely followed by weather-based models (R²=۰.۹۹۳۷). The outcomes supported the ability to provide highly precise and scalable decision outputs for production forecasting. This leads to an improved, yet practical solution to PV forecasting with relevant potential for contemporary engineering applications.

کلیدواژه ها:

Forecasting ، Hybrid Modeling ، feature engineering ، Gradient boosting ، Model Generalization