Next Article in Journal
Review of Hybrid Methods for the Characterization of Seismic Hazard in Central America
Previous Article in Journal
Vine Volume Estimation from UAV Photogrammetry and Imagery Processing
 
 
Font Type:
Arial Georgia Verdana
Font Size:
Aa Aa Aa
Line Spacing:
Column Width:
Background:
Proceeding Paper

Wind Influence on the Spatiotemporal Forecast of Global Horizontal Irradiance †

by
Llinet Benavides Cesar
*,
Miguel Ángel Manso Callejo
and
Calimanut-Ionut Cira
Departamento de Ingeniería Topográfica y Cartográfica, E.T.S.I. en Topografía, Geodesia y Cartografía, Universidad Politécnica de Madrid, C/ Mercator, 2, 28031 Madrid, Spain
*
Author to whom correspondence should be addressed.
Presented at the IV Conference on Geomatics Engineering, Madrid, Spain, 6–7 July 2023.
Environ. Sci. Proc. 2023, 28(1), 5; https://doi.org/10.3390/environsciproc2023028005
Published: 19 December 2023
(This article belongs to the Proceedings of IV Conference on Geomatics Engineering)

Abstract

:
The connection between solar irradiance and wind is a topic of interest in the field of renewable energies, as wind data have proven to be effective predictors of solar energy, being indicators of cloud movement and atmospheric conditions. This study focuses the use of decision tree-based algorithms (random forest, XGBoost and light gradient boosting machine, and LightGBM) to analyse the impact of the meridional and zonal wind components as input variables. In the study, past observations of neighbours were included as predictors to include a spatiotemporal analysis. The studied models were trained on the open, well-established OIH dataset (containing data from Oahu Island, Hawaii, located in the United States of America) featuring predominantly northeasterly winds. In the post-training analysis, it was found that the inclusion of the wind components resulted in a mean improvement of approximately 1% in the forecast skill (FS) score for all models, with the XGBoost model being the best performing model (with a 27.63% FS score).

1. Introduction

Many factors (such as cloud cover, atmospheric conditions, and the terrain characteristics) affect the amount of solar radiation that reaches the surface of the Earth. In research related to global horizontal irradiance (GHI) forecast with machine learning (ML), popular techniques used to optimise the models are the spatiotemporal approach that can capture climate conditions at the point of prediction, or the use of meteorological explanatory variables to include weather patterns. In this regard, wind data were used to train a set of autoregressive models for solar radiation forecast, and a higher forecast skill score was obtained using the dominant wind direction in the region [1]. Mukhoty et al. [2] used wind data to sequence deep learning models for solar radiation forecast and concluded that features such as wind direction, wind speed, and the GHI of neighbouring locations improved the prediction accuracy of the models.
Considering the impact of the wind direction and wind speed variables, we aim to study the impact of using its zonal and meridional components in solar radiation prediction by introducing the mentioned wind components as input variables and evaluating their influence on the predictive capability of the model implementations. To meet this objective, a spatiotemporal approach that includes lags (values at prior time steps) from neighbouring stations will be used to train and evaluate three popular tree-based ML models, namely (1) random forest, or RF, (2) gradient boosting trees, or GBT, and (3) light gradient boosting machine, or LightGBM.
The remainder of the document is structured as follows. Section 2 describes the data, the models studied, and methodology applied. Section 3 presents the results obtained and the corresponding discussion, while Section 4 draws the conclusions of the study.

2. Materials and Methods

For the study, publicly available data provided by the National Renewable Energy Laboratory, namely the OIH dataset [3], were selected. The dataset is formed by seventeen irradiance sensors (named from here on, measuring locations) located on the Oahu Island of Hawaii, with one square kilometre of spatial resolution (as shown in Figure 1) and provides a high density for the analysis to be carried out. The time resolution for measuring the GHI is one second (from March 2010 to October 2011). The data were measured in a region with known prevailing winds from the east and predominantly the northeast.
In the initial analysis of the data, the information from the ap3 measuring location was removed due to missing data; only the information from the remaining sixteen measuring locations was used. As part of the pre-processing of the data, the data were resampled to ten seconds; the night hours were removed based on a solar elevation angle with a threshold of 5° (to keep sunrise and sunset information). The GHI time series was normalised using the clearness index (calculated with k(t) = GHI/I0, where I0 is the extraterrestrial irradiance on the horizontal surface, at a t moment).
As for the ML models considered, three popular algorithms were selected as follows: (1) the RF regression model [4] forms a combination of predictor trees, the prediction delivered being formed by taking the mean of k trees; (2) the XGBoost model [5] is the implementation of the gradient boosting decision tree model that creates the trees sequentially, allowing learning from errors in the previous step (unlike RF, which creates the trees in parallel); and (3) the LightGBM model [6] is an optimisation of the XGBoost model that speeds up the training time by almost twenty times with similar accuracy levels. Therefore, the RF model is an example of a bagging ensemble model, while XGBoost is a boosting ensemble model. LightGBM is an optimisation of the XGBoost model with almost the same degree of certainty in its predictions.
As for the experimental design, an analysis of the influence of including the zonal component and meridional component was performed. The zonal component, referred to as u , is the horizontal velocity component along a circle of latitude in a west to east direction, and the meridional component, referred to as v , is the horizontal velocity component along a meridian from south to north. The problem can be defined as follows: multiple GHI series G 1 , , G d , where d is the number of irradiance measuring locations, and each G represents a time series from specific measuring locations; the model inputs are registries from the OIH dataset as G p t l , u , v ;   l   represents the number of lags, l 1 , , n , p is the number of series, p 1 , , d , u is the zonal component, v represents the meridional component, while G k t is the target of the model for the k t h measuring location.
The considered models were implemented in the scikit-learn library [7] and trained using the “GridSearch cross validation” function. During training, the hyperparameters optimised for XGBoost and LightGBM were the learning rate, l r ; l r 0.001 ,   0.01 ,   0.1 ,   0.15 ,   0.3 ,   0.45 , and the maximum depth, m d ; m d 5 ,   10 ,   15 ,   20 ,   25 , with the optimised l r being 0.1 and 10 and the optimised m d being 0.1 and 15 for XGBoost and LightGBM, respectively. The considered hyperparameters for RF were the minimum number of leaves, n l , and the number of estimators, n e ; n l , n e 100 ,   150 ,   200 ,   250 ,   300 , with the optimised n l , n e values being 0.01 and 150, respectively. All models were trained using the spatiotemporal approach (considering the lags of the neighbours) in two training scenarios (where the wind components are provided, and vice versa). The performance of the models trained with and without the inclusion of wind variables was evaluated with popular performance metrics applied to assess the accuracy of the regression models, namely, the mean absolute error (MAE), the root mean squared error (RMSE), and the forecast skill (FS) score.

3. Results and Discussion

We start with an analysis of the variability being captured and then show the results obtained by the three algorithms. In Figure 2, an example of the spatial and temporal context of the solar variability shows that, when two locations are taken on a given cloudy day, the observed GHI values can vary greatly with changes in the sun’s geometry and cloud movement.
A comparison of the evaluated metrics obtained by the considered ML models is presented in Table 1. It can be observed that all models evaluated improved their FS score with the inclusion of wind components. XGBoost was the best performer, with an average FS score of 27.63% for all measuring locations. LightGBM achieved an average FS score of 26.50%, while RF obtained an average FS value of 12.07%.
As found in Figure 3, the measuring locations with the highest FS scores are those that have the largest number of neighbours in the direction of the wind. In the ap7, ap4, and ap5 measuring locations, wind variables had less influence on the outcome and featured the lowest FS values. In the dh3, dh4, dh6, and dh10 measuring locations, the results were further improved by the introduction of wind components and displayed the best FS scores. The experiments demonstrate that the inclusion of wind components can improve the prediction results.

4. Conclusions

In this work, the influence of introducing the zonal and meridional wind components as input variables for the ML models was analysed. The study employed three tree-based algorithms for solar radiation forecast, using on OIH dataset trained using a spatiotemporal approach by including lags from neighbouring measuring stations in scenarios where wind and no wind data were used as input. The results demonstrate that using the wind components as input variables delivered an average improvement of approximately 1% in the FS value over the training scenario without wind data.
This implies that, although wind data are not highly correlated with solar radiation, they can indicate cloud movement and other atmospheric conditions and become an effective predictor of solar radiation. For future research, (1) fewer characteristics could be studied by including only the lags of neighbours in the predominant wind direction, and (2) more complex models such as those provided by deep learning implementation should be considered.

Author Contributions

Conceptualization, L.B.C. and M.Á.M.C.; methodology, L.B.C. and M.Á.M.C.; software, L.B.C.; validation, L.B.C., M.Á.M.C. and C.-I.C.; formal analysis, L.B.C.; investigation, L.B.C.; resources, M.Á.M.C. and C.-I.C.; data curation, L.B.C.; writing—original draft preparation, L.B.C.; writing—review and editing, L.B.C., M.Á.M.C. and C.-I.C.; visualization, L.B.C.; supervision, M.Á.M.C. and C.-I.C. All authors have read and agreed to the published version of the manuscript.

Funding

This research received no external funding.

Institutional Review Board Statement

Not applicable.

Informed Consent Statement

Not applicable.

Data Availability Statement

This work was based on a public dataset provided by the National Renewable Energy Laboratory. The data are freely available at http://www.nrel.gov/midc/oahu_archive/ (access on 12 December 2023).

Conflicts of Interest

The authors declare no conflict of interest.

References

  1. Amaro e Silva, R.; Haupt, S.E.; Brito, M.C. A regime-based approach for integrating wind information in spatio-temporal solar forecasting models. J. Renew. Sustain. Energy 2019, 11, 056102. [Google Scholar] [CrossRef]
  2. Mukhoty, B.P.; Maurya, V.; Shukla, S.K. Sequence to sequence deep learning models for solar irradiation forecasting. In Proceedings of the 2019 IEEE Milan PowerTech, Milan, Italy, 23–27 June 2019; IEEE: New York, NY, USA, 2019. [Google Scholar] [CrossRef]
  3. Segupta, M.; Andreas, A. Oahu Solar Measurement Grid (1-Year Archive): 1-Second Solar Irradiance; NREL Report No. DA-5500-56506; National Renewable Energy Laboratory: Oahu, HI, USA, 2010. [CrossRef]
  4. Breiman, L. Random Forests. Mach Learn 2001, 45, 5–32. [Google Scholar] [CrossRef]
  5. Chen, T.; Guestrin, C. XGBoost: A Scalable Tree Boosting System. In Proceedings of the KDD ’16: 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, San Francisco, CA, USA, 13–17 August 2016; Association for Computing Machinery: New York, NY, USA, 2016; pp. 785–794. [Google Scholar] [CrossRef]
  6. Ke, G.; Meng, Q.; Finley, T.; Wang, T.; Chen, W.; Ma, W.; Ye, Q.; Liu, T.Y. LightGBM: A Highly Efficient Gradient Boosting Decision Tree. In NIPS’17: Proceedings of the 31st International Conference on Neural Information Processing System, Long Beach, CA, USA, 4–9 December 2017; Curran Associates Inc.: Red Hook, NY, USA, 2017; pp. 3149–3157. [Google Scholar]
  7. Pedregosa, F.; Varoquaux, G.; Gramfort, A.; Michel, V.; Thirion, B.; Grisel, O.; Blondel, M.; Prettenhofer, P.; Weiss, R.; Dubourg, V.; et al. Scikit-learn: Machine Learning in Python. J. Mach. Learn. Res. 2011, 12, 2825–2830. [Google Scholar]
Figure 1. Spatial distribution of irradiance measuring locations (from the Oahu solar measurement grid). Note: the red arrow indicates the prevailing wind direction.
Figure 1. Spatial distribution of irradiance measuring locations (from the Oahu solar measurement grid). Note: the red arrow indicates the prevailing wind direction.
Environsciproc 28 00005 g001
Figure 2. GHI values obtained from the measuring locations ap7 (blue) and dh5 (orange) on a cloudy day (15 August 2010) for (a) the complete day, and (b) a thirty-minutes interval (starting at 13:00 hour). Note: Figure 2b shows drops in the GHI values produced by the clouds, which can be observed minutes later in another measuring location found in the direction of the prevailing wind.
Figure 2. GHI values obtained from the measuring locations ap7 (blue) and dh5 (orange) on a cloudy day (15 August 2010) for (a) the complete day, and (b) a thirty-minutes interval (starting at 13:00 hour). Note: Figure 2b shows drops in the GHI values produced by the clouds, which can be observed minutes later in another measuring location found in the direction of the prevailing wind.
Environsciproc 28 00005 g002
Figure 3. Forecast skill score of the XGBoost model for all the measuring locations (with and without wind information). Note: the model was trained using the spatiotemporal approach.
Figure 3. Forecast skill score of the XGBoost model for all the measuring locations (with and without wind information). Note: the model was trained using the spatiotemporal approach.
Environsciproc 28 00005 g003
Table 1. Mean performance metrics obtained by the models trained using the spatiotemporal approach for all measuring locations (grouped by the inclusion of wind data information).
Table 1. Mean performance metrics obtained by the models trained using the spatiotemporal approach for all measuring locations (grouped by the inclusion of wind data information).
Wind Variables ConsideredNo Wind Variables Considered
ModelMAE
(W/m2)
RMSE
(W/m2)
FS Score
(%)
MAE
(W/m2)
RMSE
(W/m2)
FS Score
(%)
RF27.4670.1312.0729.1971.3210.57
LightGBM22.0358.6026.5023.0559.4125.49
XGBoost20.6357.7027.6321.2058.3126.87
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.

Share and Cite

MDPI and ACS Style

Cesar, L.B.; Callejo, M.Á.M.; Cira, C.-I. Wind Influence on the Spatiotemporal Forecast of Global Horizontal Irradiance. Environ. Sci. Proc. 2023, 28, 5. https://doi.org/10.3390/environsciproc2023028005

AMA Style

Cesar LB, Callejo MÁM, Cira C-I. Wind Influence on the Spatiotemporal Forecast of Global Horizontal Irradiance. Environmental Sciences Proceedings. 2023; 28(1):5. https://doi.org/10.3390/environsciproc2023028005

Chicago/Turabian Style

Cesar, Llinet Benavides, Miguel Ángel Manso Callejo, and Calimanut-Ionut Cira. 2023. "Wind Influence on the Spatiotemporal Forecast of Global Horizontal Irradiance" Environmental Sciences Proceedings 28, no. 1: 5. https://doi.org/10.3390/environsciproc2023028005

Article Metrics

Back to TopTop