Interval Prediction Method for Solar Radiation Based on Kernel Density Estimation and Machine Learning
Open Access
- 1 January 2022
- journal article
- research article
- Published by Wiley in Complexity
- Vol. 2022 (1)
Abstract
Precise global solar radiation (GSR) data are indispensable to the design, planning, operation, and management of solar radiation utilization equipment. Some examples prove that the uncertainty of the prediction of solar radiation provides more value than deterministic ones in the management of power systems. This study appraises the potential of random forest (RF), V‐support vector regression (V‐SVR), and a resilient backpropagation artificial neural network (Rprop‐ANN) for daily global solar radiation (DGSR) point prediction from average relative humidity (RHU), daily average temperature (AT), and daily sunshine duration (SD). To acquire more accurate predictions of DGSR and examine the influence of historical DGSR on the performance of point prediction models, two different model inputs are considered: (1) three meteorological variables and (2) the lags of DGSR and three meteorological variables. Then, two interval prediction methods are developed by introducing the KDE to out‐of‐bag (OOB), introducing kernel density estimation (KDE) to split conformal (SC) based on the three machine learning models. The two methods for interval prediction are denoted as OOB‐KDE and SC‐KDE. The mean absolute error (MAE), mean relative error (MRE), and Kendall rank correlation (Kendall) are used to assess the point prediction models. The performance of interval prediction methods is evaluated by the prediction interval coverage probability (PICP), prediction interval normalized average width (PINAW), and coverage width criteria (CWC). The following conclusions are drawn from this study. First, the V‐SVR model performs best with the lowest mean absolute error (MAE) of 0.016 and mean relative error (MRE) of 0.001. Second, the lags of DGSR improve the prediction accuracy by about 30%. Third, the OOB‐KDE and SC‐KDE methods improved the quality of the prediction interval (PI). OOB‐KDE improved CWC by 81%, and SC‐KDE improved CWC by 99.99%. Fourth, the best interval prediction result is obtained using the SC‐KDE method using the V‐SVR model. The average difference between its PICP and prediction interval nominal coverage (PINC) is only 3% of the PINC, and its PINAW is less than 0.007.Keywords
Funding Information
- National Natural Science Foundation of China (12171328, 12031016)
- Capital Normal University (00719530012012, 00719530012010)
This publication has 43 references indexed in Scilit:
- Comparison of ANN and MLR models for estimating solar radiation in Turkey using NOAA/AVHRR dataAdvances In Space Research, 2013
- Applying Rprop Neural Network for the Prediction of the Mobile Station LocationSensors, 2011
- Predicting solar radiation at high resolutions: A comparison of time series forecastsSolar Energy, 2009
- On boundary correction in kernel density estimationStatistical Methodology, 2005
- Multidimensional Density Estimation2005
- Statistical Modeling: The Two Cultures (with comments and a rejoinder by the author)Statistical Science, 2001
- Bagging predictorsMachine Learning, 1996
- Learning representations by back-propagating errorsNature, 1986
- An alternative method of cross-validation for the smoothing of density estimatesBiometrika, 1984
- Non-Parametric Estimation of a Multivariate Probability DensityTheory of Probability & Its Applications, 1969