Time series cross-correlation network for wind power prediction

Wind power is an indispensable part of clean energy, but due to its inherent instability, it is necessary to predict the power generation of wind turbines after a period of time accurately. Most recent approaches are based on machine learning methods, which map time series data to a high-dimensional space, follow Markov process in the time dimension, and extract time series features following the chronological order. However, due to the time-instability and highly spatially correlated nature of wind power data, the prediction methods which follow the chronological order cannot extract all features in wind power time series. This will lose the information contained in the sequence and lack spatially relevant information. This paper proposes a TSCN-LSTM model that includes a Time Series Cross-correlation Network (TSCN) and a long short term memory (LSTM) decoder to predict wind power generation after 30 minutes. TSCN-LSTM not only collaboratively encodes the neighboring area and the neighboring time to fully tap the potential spatiotemporal correlation by TSCN, but also uses the LSTM decoder to enhance the timing relationship and to prevent the loss of timing information. At the same time, a data preprocessing method is proposed to enhance the spatial representation of the data. It makes that TSCN-LSTM can integrate the temporal and spatial feature extraction process to enhance the semantic expression of features. By establishing extensive interconnections between data in time dimension and space dimension, multiple types of cross-correlation information are generated which enables the model to distinguish different meteorological features and discover temporal and spatial correlations. A large number of experiments show that compared with the current leading SVR, LSTM, LSTM-EFG, GRU, Bi-LSTM and SATCN-LSTM methods, the wind power prediction mean square error is reduced by an average of 13.06%, 12.44%, 4.05%, 5.11%, 6.94% and 8.76%, respectively.