A selective memory attention mechanism for chaotic wind speed time series prediction with auxiliary variable

Wind speed prediction is crucial for enhancing wind energy utilization and optimizing grid integration of wind power. Its chaotic nature and the lack of correlated variables make accurate prediction difficult. Most studies rely solely on past wind speed, limiting accuracy improvements. While wind power is highly correlated with wind speed, this correlation is reversely causal. The key challenge is effectively leveraging this reverse causality between wind power and wind speed to enhance prediction precision. This study proposed SMAMnet to address the challenge mentioned, a model that establishes its backbone network via proposed new attention mechanism. The convolution operation is employed to restructure features, besides, the frequency-domain transformation and selective state space model (SSM) serves for attention weights. The novelty of SMAMnet is characterized by the development of an adaptive frequency-domain selected attention weight operator to adaptively parse meaningful information in different frequency domain intervals. Taking 15 min and 1-hour mean absolute error as the standard, the actual wind speed prediction error is reduced by 68% and 49% compared with the classic LSTM algorithm. The feasibility of mining reverse causality to improve prediction accuracy was verified.