Constructing Probabilistic Load Forecast from Multiple Point Forecasts: A Bootstrap Based Approach

Abstract

Probabilistic load forecast presents more information on the possible deviation of forecast than the point forecast. There are sufficient regression models that can make point forecasts. An intuitive question can be raised: Is there a way to combine the point forecasts to construct a probability or interval forecast? In this paper, a bootstrap based ensemble approach is put forward to construct forecast intervals from multiple point forecasts. Specifically, multiple point forecasting models are first trained based on the bootstrap sampled training datasets and different forecasting models. Then, bootstrap is applied again to the multiple point forecasts. Finally, the quantiles are estimated according to the distribution of the sampled point forecasts. Two common machine learning methods, random forest (RF) and gradient boosting regression tree (GBRT), are combined to test the feasibility of the proposed forecasting framework. Compared with quantile RF (Q-RF) and quantile GBRT (Q-GBRT), numerical experiments demonstrate its advantage over Q-RF and Q-GBRT.