Very Short-Term Current and Load Forecasting for Distribution Systems in Data Constrained Situations

Load and current forecasting are useful tools for distribution systems (DSs) real-time monitoring and operation. Many techniques have been successfully applied to this task with various different time horizons. For real-time operation, forecasting algorithms must provide fast answers and be able to deal with high granularity and data availability that may be much smaller than in ideal conditions. This paper presents a methodology that combines renowned techniques, gradient boosting and persistence, into a predictor that can adjust itself to real-time changes in DSs and requires very little data to be trained. This method is tested using only 21 days as training dataset to predict the behavior of a portion of a real DSs with average errors lower than 5%.