A Wind Power Forecasting Method Considering Shape and Time Loss

To improve the accuracy of wind power time series with strong nonlinearity and non-stationarity, this paper proposed a wind power forecasting (WPF) method based on deep learning, which considers shape and time loss. First, A two-layer Gated Recurrent Unit (GRU) is used to construct a WPF model. Then, this paper uses differentiable dynamic time warping (DTW) to construct shape and time loss (STL) to improve the loss function of the GRU. STL decouples the time series in shape and time and provides their loss metrics. Finally, the test results on the wind power data of the Global Energy Forecasting Competition 2014 (GEFCom2014) show that compared with the traditional mean squared error (MSE) loss, the STL has a higher prediction accuracy on abrupt and non-stationary time series.