Prediction Interval Construction for Electric Load and Wind Power via Machine Learning

Abstract

Accurate and reliable forecasting methods are essential to smart grid operation and planning. However, due to intrinsic uncertainties such as non-deterministic electricity loads and intermittent power generation from renewable energy sources, traditional point forecasting methods can hardly be accurate as only one predictive value is generated at each time step. To assess multifarious uncertainties, probabilistic forecasting, which can be in the form of prediction intervals (PIs), is preferred for forecasting tasks in future smart grid. This paper proposes a new ensemble machine learning approach to enhance both the reliability and sharpness of PIs. In the proposed approach, a novel framework incorporating point forecast and stochastic gradient boosted quantile regression is created to identify and modify illogical PIs. Furthermore, an explicit procedure for implementing the proposed approach, and an empirical parameter tuning method are proposed. The performance of the proposed method is validated for two forecasting tasks in a smart grid, namely load forecasting and wind power forecasting. Compared with benchmark models, the proposed approach is more robust, and exhibits a significantly enhanced performance, as measured by multiple PI evaluation indexes.