Multi-layer perceptron hybrid model integrated with the firefly optimizer algorithm for windspeed prediction of target site using a limited set of neighboring reference station data

Long-term windspeed prediction is crucial for establishing the viability of wind as a clean energy option, including the selection of wind farm locations, feasibility studies on energy potential and the operation of wind energy conversion systems with minimal investment risk. To deliver this vital societal need, data-inexpensive artificial intelligence models relying on historical inputs can be a useful scientific contrivance by energy analysts, engineers and climate-policy advocates. In this paper, a novel approach is adopted to construct a multilayer perceptron (MLP) hybrid model integrated with the Firefly Optimizer algorithm (MLP-FFA) trained with a limited set of historical (monthly) data (2004–2014) for a group of neighboring stations to predict windspeed at target sites in north-west Iran. Subsequently, the MLP-FFA model is developed to minimize the error rate of the resulting hybrid model and applied at each of the eight target sites one-by-one (namely: Tabriz, Jolfa, Sarab, Marand, Sahand, Kaleybar, Maraghe and Mianeh) such that the seven neighboring (reference) sites are used for training and the remainder eighth site for testing purposes. To ascertain conclusive results, the hybrid model's ability to predict windspeed at each target site is cross-validated with the MLP model without the FFA optimizer and the statistical performance is benchmarked with root mean square error (RMSE), mean absolute error (MAE), Nash-Sutcliffe efficiency (ENS), Willmott's Index (d) and the Legates and McCabes Index (E1), including relative errors. For all eight target sites, the testing performance of the MLP-FFA model is found to be significantly superior than the classical MLP, resulting in lower values of the RMSE (0.202–0.50 ms−1 relative to 0.236–0.664 ms−1) and larger values of ENS, d and E1 (0.686–0.953 vs. 0.529–0.936, 0.874–0.976 vs. 0.783–0.966, 0.417–0.800 vs. 0.303–0.748). Despite a more accurate performance of hybrid models tested at each target site, the preciseness registered a distinct geographic signature with the least accurate result (for Kaleybar) and the most accurate result (for Jolfa). To accord with this result, we conclude that the utilization of the FFA as an add-in optimizer in a hybrid data-intelligent model leads to a significant improvement in the predictive accuracy, presumably due to the optimal weights attained in the hidden layer that allows a more robust feature extraction process. Accordingly, we establish that the hybrid MLP-FFA model can be explored further in a problem of long-term windspeed prediction with reference station input data, and feasibility studies on wind energy investments in data-scarce regions where a limited set of neighboring reference site data can be employed to forecast the target site windspeed.