A novel EMD and causal convolutional network integrated with Transformer for ultra short-term wind power forecasting

Abstract

Accurate wind power forecasting can enhance the safety, stability, economy and controllability of the power system. Traditional physical methods and statistical methods are easily affected by data quality and extraction methods in wind power forecasting. The commonly used recursive neural network method may have memory decline phenomenon in wind power forecasting and does not support parallel calculation, thus limiting the forecasting accuracy. To solve the above problems, in this paper, a wind power forecasting method based on EMD-CCTransformer is proposed. The network model is based on an encoder–decoder structure, where the encoder is used to parse historical wind power sequences, the decoder generates future wind power, and the encoder and decoder are connected using an attention mechanism. In this method, the EMD algorithm is used to decompose the wind power series and obtain the changes of power signals in different time scales, which improves the ability of Transformer to maintain long-term information. Meanwhile, in view of the partial unknowability of the Transformer model, the convolutional attention mechanism is introduced to replace the dot product attention mechanism to form the CCTransformer model, which further improves the forecasting accuracy. We use large-scale wind power data (annual data of the year of 2020 ) to train and test the proposed model. Experimental results show that compared with commonly used wind power forecasting methods, the forecasting error of the proposed EMD-CCTransformer forecasting method in this paper is lower and its model training time is further shortened.