- This work can be found in [Renewable Energy].
📍 This work proposes a wind power prediction model, in which the proposed model uses self-attention to capture the long-range relationship and uses convolutional layers to learn the local temporal interactions of the variables in the time-series data.
📍 Compared to deep learning sequence models, such as recurrent neural network (RNN), long short-term memory (LSTM), and gated recurrent unit (GRU), the proposed method can simultaneously consider global and local information.
📍 It comprises a feedforward module, an attention module, and a convolution module. Our proposed model is inspired by the Transformer [1].
📍 The proposed model mainly uses self-attention to capture long-range relationships by repeatedly stacking self-attention modules, but the local temporal correlation is also important in time-series forecasting.
📍 Thus, we also use the convolution module to enhance the acquisition of local features.
The data used in the wind power prediction task belong to time-series data, which means that a series of data points are collected over time. We introduce the notation for time series data at different time steps as x1, x2, . . . , xT to denote time series data of length T, each time steps has
we use several methods that are often used in wind energy forecasting as comparison methods.
*The result of the forecast performance shows that the prediction values are close to the true values, indicating that the proposed model has successfully captured the information on the short-term information.