Wind Power Forecasting Using Hybrid ANN–NWP Models

In an era of growing wind energy penetration, accurate wind power forecasts are increasingly important. Artificial Neural Networks (ANNs) typically minimize mean squared error, which is not well suited to wind power forecasts because their error distribution is non-Gaussian. This study presents a two step methodology for 7-day-ahead wind power forecasts based on experiments with different methods of a) correcting raw model wind speed forecasts, and b) producing wind power forecasts based on these wind speed forecasts. Wind speed correction methods include a bias correction algorithm and two ANNs: a Multi-Layer Perceptron (MLP) and a Long-Short Term Memory Network (LSTM). Wind power conversion methods include a look-up table (LUT), a polynomial fit of the power curve, and the same two ANNs: MLP and LSTM. For wind speed correction, ANNs outperformed bias correction by an average 14.79% (15.69%) for MAE (RMSE). For wind power conversion, optimal performance was found fitting a polynomial power curve. This, in conjunction with ANN-corrected wind speeds, produces an average 10.25% (13.16%) improvement in wind power MAE (RMSE) over the reference combination of bias correction and a LUT. At one of the test wind farms, an improvement of up to 15.73% (18.49%) for the same metrics was observed. Methods using an ANN only in the wind speed correction step produced better results than those using an ANN only in the power conversion step. This highlights the fact that a two step approach of wind speed correction followed by wind power conversion is optimal when dealing with ANN models. No advantage was gained by applying ANNs in both steps. Furthermore, in either step the simpler MLP outperformed the more computationally expensive LSTM.