@article{Alrashidi2026, 
author = {Musaed Alrashidi},
title = {A hybrid VMD-PermEn-DBSCAN-ICEEMDAN deep learning framework for ultra-short-term wind speed prediction},
year = {2026},
journal = {AIMS Mathematics},
volume = {11},
number = {5},
pages = {12548-12579},
keywords = {time series forecast, wind speed, deep learning, signal decomposition methods, clustering, energy systems},
url = {https://www.sciopen.com/article/10.3934/math.2026516},
doi = {10.3934/math.2026516},
abstract = {Accurate ultra-short-term wind speed forecasting is essential for the reliable integration of wind energy into power grids; nevertheless, it is challenging due to the non-linearity and non-stationarity of wind signals. Therefore, this research introduces a novel multi-phase hybrid framework—VMD-PermEn-DBSCAN-ICEEMDAN—aimed at improving prediction accuracy through the systematic refinement of complex signal components. The process initiates with variational mode decomposition (VMD) to decompose raw wind speed data into intrinsic mode functions (IMFs). Permutation entropy (PermEn) is employed for feature extraction to address the complexity of these components, followed by density-based spatial clustering of applications with noise (DBSCAN) clustering to categorize IMFs exhibiting analogous dynamic patterns. A secondary decomposition phase employing enhanced complete ensemble empirical mode decomposition with adaptive noise (ICEEMDAN) is aimed primarily at high-frequency clusters to reveal hidden fluctuations. These refined features serve as inputs for advanced deep learning models, such as long short-term memory (LSTM) networks, gated recurrent units (GRU), and their hybrid configurations. The framework was assessed utilizing wind speed data from Riyadh, Saudi Arabia, gathered at 5-minute intervals. Experimental findings indicated that the LSTM-GRU hybrid model consistently surpasses independent architectures and conventional machine learning methods, including artificial neural networks, support vector machines, and decision trees. The proposed framework attained a remarkable mean squared error of 0.00081 m2/s2 and a coefficient of determination of 0.99954 for 5-minute forecasts. In addition, the study examined the effects of input lag lengths and forecasting resolutions of up to one hour, validating the model's durability and exceptional performance in ultra-short-term scenarios. The results underscore the effectiveness of integrating adaptive signal decomposition, intelligent clustering, and deep learning for accurate wind speed prediction, offering a dependable resource for energy management and grid stability.}
}