Hybrid convolutional Bi-LSTM autoencoder framework for short-term wind speed prediction

Accurate wind speed prediction is essential for optimal operation and planning. The unstable and stochastic nature of the wind makes the task complicated and challenging. As a result, a hybrid approach is implemented to enhance prediction accuracy and to overcome the difficulties and challenges in uncertainty modelling. Encoder and decoder are the two parts of the proposed hybrid model. In this study, a one-dimensional convolutional neural network (1D-CNN) is used as the encoder, and a bidirectional long short term memory network (Bi-LSTM) is used as the decoder. Encoder extracts the important characteristics and forms a latent representation. Then, wind speed is predicted by the decoder network by interpreting the characteristics of the encoded representation. The hybrid approach is validated using several regular and widely used benchmark forecasting models to assess and examine its prediction performance. The prediction results using the real-time dataset obtained from a wind measuring station in Idalia, Colorado are used for performance evaluation. The performance validation analysis showed that the proposed hybrid approach has an improvement of 42% over the reference approaches.

Automated summary

Accurate wind speed prediction is crucial for optimal operation and planning, but the unstable and stochastic nature of wind poses challenges. This study proposes a hybrid approach that utilizes an encoder and a decoder to improve prediction accuracy and address uncertainty modeling difficulties. The encoder, a one-dimensional convolutional neural network, extracts important characteristics and forms a latent representation, while the decoder, a bidirectional long short term memory network, interprets these characteristics to predict wind speed. The hybrid approach is validated and outperforms benchmark models by 42% in real-time wind data from a measuring station in Colorado.