Evaluation of regression and neural network models for solar forecasting over different short-term horizons

Alireza Inanlouganji, T Agami Reddy, Srinivas Katipamula

Research output: Contribution to journalArticle

2 Citations (Scopus)

Abstract

Forecasting solar irradiation has acquired immense importance in view of the exponential increase in the number of solar photovoltaic (PV) system installations. In this article, analyses results involving statistical and machine-learning techniques to predict solar irradiation for different forecasting horizons are reported. Yearlong typical meteorological year 3 (TMY3) datasets from three cities in the United States with different climatic conditions have been used in this analysis. A simple forecast approach that assumes consecutive days to be identical serves as a baseline model to compare forecasting alternatives. To account for seasonal variability and to capture short-term fluctuations, different variants of the lagged moving average (LMX) model with cloud cover as the input variable are evaluated. Finally, the proposed LMX model is evaluated against an artificial neural network (ANN) model. How the one-hour and 24-hour models can be used in conjunction to predict different short-term rolling horizons is discussed, and this joint application is illustrated for a four-hour rolling horizon forecast scheme. Finally, the effect of using predicted cloud cover values, instead of measured ones, on the accuracy of the models is assessed. Results show that LMX models do not degrade in forecast accuracy if models are trained with the forecast cloud cover data.

Original languageEnglish (US)
Pages (from-to)1-10
Number of pages10
JournalScience and Technology for the Built Environment
DOIs
StateAccepted/In press - May 24 2018

Fingerprint

Neural networks
Irradiation
Learning systems

ASJC Scopus subject areas

  • Environmental Engineering
  • Building and Construction
  • Fluid Flow and Transfer Processes

Cite this

Evaluation of regression and neural network models for solar forecasting over different short-term horizons. / Inanlouganji, Alireza; Reddy, T Agami; Katipamula, Srinivas.

In: Science and Technology for the Built Environment, 24.05.2018, p. 1-10.

Research output: Contribution to journalArticle

@article{3920456f789e4e4a8d3f5e6eaaa8c787,
title = "Evaluation of regression and neural network models for solar forecasting over different short-term horizons",
abstract = "Forecasting solar irradiation has acquired immense importance in view of the exponential increase in the number of solar photovoltaic (PV) system installations. In this article, analyses results involving statistical and machine-learning techniques to predict solar irradiation for different forecasting horizons are reported. Yearlong typical meteorological year 3 (TMY3) datasets from three cities in the United States with different climatic conditions have been used in this analysis. A simple forecast approach that assumes consecutive days to be identical serves as a baseline model to compare forecasting alternatives. To account for seasonal variability and to capture short-term fluctuations, different variants of the lagged moving average (LMX) model with cloud cover as the input variable are evaluated. Finally, the proposed LMX model is evaluated against an artificial neural network (ANN) model. How the one-hour and 24-hour models can be used in conjunction to predict different short-term rolling horizons is discussed, and this joint application is illustrated for a four-hour rolling horizon forecast scheme. Finally, the effect of using predicted cloud cover values, instead of measured ones, on the accuracy of the models is assessed. Results show that LMX models do not degrade in forecast accuracy if models are trained with the forecast cloud cover data.",
author = "Alireza Inanlouganji and Reddy, {T Agami} and Srinivas Katipamula",
year = "2018",
month = "5",
day = "24",
doi = "10.1080/23744731.2018.1464348",
language = "English (US)",
pages = "1--10",
journal = "Science and Technology for the Built Environment",
issn = "2374-4731",
publisher = "Taylor and Francis Ltd.",

}

TY - JOUR

T1 - Evaluation of regression and neural network models for solar forecasting over different short-term horizons

AU - Inanlouganji, Alireza

AU - Reddy, T Agami

AU - Katipamula, Srinivas

PY - 2018/5/24

Y1 - 2018/5/24

N2 - Forecasting solar irradiation has acquired immense importance in view of the exponential increase in the number of solar photovoltaic (PV) system installations. In this article, analyses results involving statistical and machine-learning techniques to predict solar irradiation for different forecasting horizons are reported. Yearlong typical meteorological year 3 (TMY3) datasets from three cities in the United States with different climatic conditions have been used in this analysis. A simple forecast approach that assumes consecutive days to be identical serves as a baseline model to compare forecasting alternatives. To account for seasonal variability and to capture short-term fluctuations, different variants of the lagged moving average (LMX) model with cloud cover as the input variable are evaluated. Finally, the proposed LMX model is evaluated against an artificial neural network (ANN) model. How the one-hour and 24-hour models can be used in conjunction to predict different short-term rolling horizons is discussed, and this joint application is illustrated for a four-hour rolling horizon forecast scheme. Finally, the effect of using predicted cloud cover values, instead of measured ones, on the accuracy of the models is assessed. Results show that LMX models do not degrade in forecast accuracy if models are trained with the forecast cloud cover data.

AB - Forecasting solar irradiation has acquired immense importance in view of the exponential increase in the number of solar photovoltaic (PV) system installations. In this article, analyses results involving statistical and machine-learning techniques to predict solar irradiation for different forecasting horizons are reported. Yearlong typical meteorological year 3 (TMY3) datasets from three cities in the United States with different climatic conditions have been used in this analysis. A simple forecast approach that assumes consecutive days to be identical serves as a baseline model to compare forecasting alternatives. To account for seasonal variability and to capture short-term fluctuations, different variants of the lagged moving average (LMX) model with cloud cover as the input variable are evaluated. Finally, the proposed LMX model is evaluated against an artificial neural network (ANN) model. How the one-hour and 24-hour models can be used in conjunction to predict different short-term rolling horizons is discussed, and this joint application is illustrated for a four-hour rolling horizon forecast scheme. Finally, the effect of using predicted cloud cover values, instead of measured ones, on the accuracy of the models is assessed. Results show that LMX models do not degrade in forecast accuracy if models are trained with the forecast cloud cover data.

UR - http://www.scopus.com/inward/record.url?scp=85047432203&partnerID=8YFLogxK

UR - http://www.scopus.com/inward/citedby.url?scp=85047432203&partnerID=8YFLogxK

U2 - 10.1080/23744731.2018.1464348

DO - 10.1080/23744731.2018.1464348

M3 - Article

AN - SCOPUS:85047432203

SP - 1

EP - 10

JO - Science and Technology for the Built Environment

JF - Science and Technology for the Built Environment

SN - 2374-4731

ER -