A wavelet-coupled support vector machine model for forecasting global incident solar radiation using limited meteorological dataset

A wavelet-coupled support vector machine model for forecasting global incident solar radiation using limited meteorological dataset

A solar radiation forecasting model can be utilized is a scientific contrivance for investigating future viability of solar energy potentials. In this paper, a wavelet-coupled support vector machine (W-SVM) model was adopted to forecast global incident solar radiation based on the sunshine hours (St), minimum temperature (Tmax), maximum temperature (Tmax), windspeed (U), evaporation (E) and precipitation (P) as the predictor variables. To ascertain conclusive results, the merit of the W-SVM was benchmarked with the classical SVM model. For daily forecasting, sixteen months of data (01-March-2014 to 30-June-2015) partitioned into the train (65%) and test (35%) set for the three metropolitan stations (Brisbane City, Cairns Aero and Townsville Aero) were utilized. Data were decomposed into their wavelet sub-series by discrete wavelet transformation algorithm and summed up to create new series with one approximation and four levels of detail using Daubechies-mother wavelet. For daily forecasting, six model scenarios were formulated where the number of input was increased and the forecast was assessed by statistical metrics (correlation coefficient r; Willmott's index d; Nash-Sutcliffe coefficient ENS; peak deviation Pdv), distribution statistics and prediction errors (mean absolute error MAE; root mean square error RMSE; mean absolute percentage error MAPE; relative root mean square error RMSE). Results for daily forecasts showed that the W-SVM model outperformed the classical SVM model for optimum input combinations. A sensitivity analysis with single predictor variables yielded the best performance with St as an input, confirming that the largest contributions for forecasting solar energy is derived from the sunshine hours per day compared to the other prescribed inputs. All six inputs were required in the optimum W-SVM for Brisbane and Cairns stations to yield r = 0.928, d = 0.927, ENS = 0.858, Pdv = 1.757%, MAE = 1.819 MJ m-2 and r = 0.881, d = 0.870, ENS = 0.762, Pdv = 9.633%, MAE = 2.086 MJ m-2, respectively. However, for Townsville, the time-series of St, Tmin and Tmax and E were required in the optimum model to yield r = 0.858, d = 0.886, ENS = 0.722, Pdv = 10.282% and MAE = 2.167 MJ m-2. In terms of the relative model errors over daily forecast horizon, W-SVM model was the most accurate precise for Townsville (RRMSE = 12.568%; MAPE = 12.666%) followed by Brisbane (13.313%; 13.872%) and Cairns (14.467%; 15.675%) weather stations. A set of alternative models developed over the monthly, seasonal and annual forecast horizons verified the long-term forecasting skill, where lagged inputs of Tmax, Tmin, E, P and VP for Roma Post Office and Toowoomba Regional stations were employed. The wavelet-coupled model performed well, with r = 0.965, d = 0.964, Pdv = 2.249%, RRMSE = 5.942% and MAPE = 4.696% (Roma) and r = 0.958, d = 0.943, Pdv = 0.979%, RRMSE = 7.66% and MAPE = 6.20% (Toowoomba). Accordingly, the results conclusively ascertained the importance of wavelet-coupled SVM predictive model as a qualified stratagem for short and long-term forecasting of solar energy for assessment of solar energy prospectivity in this study region.

Permanent restricted access to published version due to publisher copyright policy.