Prediction intervals for global solar irradiation forecasting using regression trees methods

Abstract

A global horizontal irradiation prediction (from 1 hour to 6 hours) is performed using 2 persistence models (simple and "smart" ones) and 4 machine learning tools belonging to the regression trees methods family (normal, pruned, boosted and bagged). A prediction band is associated to each forecast using methodologies based on: bootstrap sampling and k-fold approach, mutual information, stationary time series process with clear sky model, quantiles estimation and cumulative distribution function. New reliability indexes (gamma index and gamma test) are built from the mean interval length (MIL) and prediction interval coverage probability (PCIP). With such methods and error metrics, good prediction bands are estimated for Ajaccio (France) with a MIL close to 113 Wh/m², a PCIP reaching 70% and a gamma index lower than 0.9.

Comparison of statistical post-processing methods for probabilistic NWP forecasts of solar radiation

“…Lastly, Voyant et al (2018) made a comparison between probabilistic forecasts using quantile regression forests and gradient boosting decision trees with past CSI observations as predictor set. They generate prediction intervals for the CSI and verify them using the gamma test, which is based on the mean interval length and prediction interval coverage probability.…”

Pattern Recognition and Its Application in Solar Radiation Forecasting

“…K-fold cross validation approach is a resampling procedure used to evaluate machine learning models over a sample data. The bootstrap is another resampling technique that generates multiple datasets by sampling from the original single dataset [13]. A combination of bootstrap sampling, k-fold cross validation, mutual information, and stationary time-series processing with clear sky model is applied in Ref.…”

“…A combination of bootstrap sampling, k-fold cross validation, mutual information, and stationary time-series processing with clear sky model is applied in Ref. [13] for solar radiation forecasting. In this method, the cumulative distribution function (CDF) and matching quintile estimation (MQE) are used to determine the prediction interval.…”

Prediction interval methodology based on fuzzy numbers and its extension to fuzzy systems and neural networks

“…Other approaches regarding the prediction intervals of renewable resources, the price of energy, and the electricity demand have been reported ( Hu, Hu, Yue, Zhang, & Hu, 2017;Li et al, 2018;Shrivastava et al, 2015Shrivastava et al, , 2016Voyant et al, 2018 ). In the works of Shrivastava et al (2016) and Shrivastava et al (2015) , methodologies were proposed based on the support vector machine (SVM) to generate the prediction intervals for wind speed and electricity costs.…”

“…The optimization was performed using a cost function that included the coverage probability, the sharpness of the interval and the average deviation of the data from the prediction interval as metrics. In the work of Voyant et al (2018) , prediction interval models of the global horizontal irradiation using regression tree methods were presented. Several prediction models were tested, including classic, pruned, bagged and boosted regression tree and classic and smart persistence models.…”