STL Enhanced Models for Mid-Term Reactive Power Forecasting

This paper presents a framework for mid-term reactive power forecasting at the transmission–distribution interface. Seasonal-trend decomposition with locally estimated scatterplot smoothing (STL-LOESS) is used as a preprocessing step to isolate short-term variations from non-linear trends and seasonal effects. Forecasting models are trained on historical data from three power transformers using both STL-processed and raw time series. The study evaluates Extreme Gradient Boosting (XGBoost), Light Gradient Boosting (LightGBM), and Long Short-Term Memory (LSTM) models, representing machine learning and deep learning techniques, across both configurations. The results indicate that errors increase with trend fluctuations. However, STL preprocessing decreases the RMSE by up to 0.30 p.u. and enhances the model performance in various transformer profiles.