Considering that wind power is proportional to the cube of the wind speed variable, which is highly random, complex power grid management tasks have arisen as a result. Wind speed prediction in the short term is crucial for load dispatch planning and load increment/decrement decisions. The chaotic intermittency of speed is often characterised by inherent linear and nonlinear patterns, as well as nonstationary behaviour; thus, it is generally difficult to predict it accurately and efficiently using a single linear or nonlinear model. In this study, wavelet transform (WT), autoregressive integrated moving average (ARIMA), extreme gradient boosting trees (XGBoost), and support vector regression (SVR) are combined to predict high-resolution short-term wind speeds obtained from three Southern African Universities Radiometric Network (SAURAN) stations: Richtersveld (RVD); Central University of Technology (CUT); and University of Pretoria (UPR). This hybrid model is termed WT-ARIMA-XGBoost-SVR. In the proposed hybrid, the ARIMA component is employed to capture linearity, while XGBoost captures nonlinearity using the wavelet decomposed subseries from the residuals as input features. Finally, the SVR model reconciles linear and nonlinear predictions. We evaluated the WT-ARIMA-XGBoost-SVR’s efficacy against ARIMA and two other hybrid models that substitute XGBoost with a light gradient boosting machine (LGB) component to form a WT-ARIMA-LGB-SVR hybrid model and a stochastic gradient boosting machine (SGB) to form a WT-ARIMA-SGB-SVR hybrid model. Based on mean absolute error (MAE), mean absolute percentage error (MAPE), root mean square error (RMSE), coefficient of determination (R2), and prediction interval normalised average width (PINAW), the proposed hybrid model provided more accurate and reliable predictions with less uncertainty for all three datasets. This study is critical for improving wind speed prediction reliability to ensure the development of effective wind power management strategies.
The global horizontal irradiance data recorded at the earth’s horizontal surface is a mixture of deterministic (extra-terrestrially) and stochastic components due to the ever-changing atmospheric conditions. The Box-Jenkins short memory stochastic models and their hybrid versions have been used successfully to forecast solar irradiance data. However, these models lack robustness and faulter for distant horizon forecasting such as more than 2 hours in the hourly case. Using a quantile regression model as both a robust benchmark and robustification model, we remedy this drawback by using long memory models and their hybrid versions in 24 hour ahead irradiance forecasting at three sites in South Africa (RVD, UPR and SUN). Results show that the inherent long memory is anti-persistent in all models but one that shows persistence. The forecasts from these long memory-based models are generally within the prediction interval of the benchmark model forecasts for two sites. Based on the RMSE, rRMSE, MAE, rMAE and the pinball loss function at the central quantile level and two extreme quantile levels, forecasts obtained through robustification via forecasts combinations with the quantile regression yield the best performance. The predictive ability of the best model before forecasts combination against the quantile regression model combined (robustified) model based Diebold-Mariano test rejected the null hypothesis of the same accuracy. As a follow up to distinguish which has a superior predictive accuracy "murphy diagrams" were used. For the RVD and UPR sites, the model based on forecast combination has the best predictive accuracy while the quantile regression based benchmark model has the best accuracy for the SUN data.
Although regression quantiles (RQs) are increasingly becoming popular, they are still playing a second fiddle role to the ordinary least squares estimator like their robust counterparts due to the perceived complexity of the robust statistical methodology. In order to make them attractive to statistical practitioners, an endeavor to studentize robust estimators has been undertaken by some researchers. This paper suggests two versions of RQs studentized residual statistics, namely, internally and externally studentized versions based on the elemental set method. The more preferred externally studentized version is compared to the one based on standardized median absolute deviation (MAD) of residuals using a well-known data set in the literature. While the MAD based outlier diagnostic seemed to be uniform and more aggressive to flagging outliers the RQ externally studentized one exhibited a dynamic pattern consistent with RQ results.
The importance of variable selection and regularization procedures in multiple regression analysis cannot be overemphasized. These procedures are adversely affected by predictor space data aberrations as well as outliers in the response space. To counter the latter, robust statistical procedures such as quantile regression which generalizes the well-known least absolute deviation procedure to all quantile levels have been proposed in the literature. Quantile regression is robust to response variable outliers but very susceptible to outliers in the predictor space (high leverage points) which may alter the eigen-structure of the predictor matrix. High leverage points that alter the eigen-structure of the predictor matrix by creating or hiding collinearity are referred to as collinearity influential points. In this paper, we suggest generalizing the penalized weighted least absolute deviation to all quantile levels, i.e., to penalized weighted quantile regression using the RIDGE, LASSO, and elastic net penalties as a remedy against collinearity influential points and high leverage points in general. To maintain robustness, we make use of very robust weights based on the computationally intensive high breakdown minimum covariance determinant. Simulations and applications to well-known data sets from the literature show an improvement in variable selection and regularization due to the robust weighting formulation.
24 July 2014 in Communications in Statistics - Theory and Methods
Regression Quantiles (RQs) (see Koenker and Bassett, 1978) can be found as optimal solutions to a Linear Programming (LP) problem. Also, these optimal solutions correspond to specific elemental regressions (ERs). On the other hand, single case ordinary least squares (OLS) leverage statistics can be expressed as weighted averages of ER ones. Using this three-tier relationship amongst RQs, ERs, and OLS leverage statistics some relationships between single case leverage statistics and ER ones are explored and deduced. We build upon these results and propose a multiple-case RQ weighted predictive leverage statistic, TJ. We do this using an ER view of the well-known leverage relationship, , by summing the ER weighted predictive leverage statistics over all ERs (RQs included) instead of over observations, i.e., . As an ad-hoc cut-off value of this statistic we make use of the analog of the Hoaglin and Welsch (1978) one, i.e., high leverage points have . So in the RQ weighted predictive leverage scenario, the cut-off value becomes , where K is the total number of ERs. We then apply this RQ high leverage diagnostic to well-known data sets in the literature. The cut-off value used generally seems too small. Some proposals of cut-off values based on some analytical bounds and a simulation study are therefore given and shown to be reasonable.