Ultra-Short-Term Wind Power Forecasting Based on Deep Embedded Clustering and LSTM

The accuracy of wind power forecasting is significantly affected by its high volatility. Clustering sequences with similar fluctuations and training models separately can improve forecasting accuracy. Therefore, this paper proposed a novel approach for ultra-short-term wind power forecasting, combining Deep Embedded Clustering (DEC) with Long Short-Term Memory (LSTM). DEC is used to cluster the wind power sequence, then the LSTM models are trained and stored separately for each cluster. Subsequently, the distance between the wind power data to be forecast and different clusters is calculated, the LSTM model corresponding to the nearest cluster is selected for forecasting, and the forecasting result is output. The results of the experiments demonstrate that compared to the LSTM and K-means-LSTM, the DEC-LSTM exhibits superior accuracy in wind power forecasting. Moreover, the proposed method also performs well in other wind farms within the wind farm cluster. Simultaneously, the Autoencoder (AE) are used in DEC for feature learning during the clustering process, and their clustering results are interpretable.