A hybrid LPMS-Transformer-ARFIMA framework for wind power forecasting with SPSO-based global hyperparameter optimization
TL;DR
An advanced hybrid framework for wind power forecasting, VMD-PSO-LPMS-Transformer-ARFIMA, with global hyperparameter optimization based on SPSO, based on a surrogate assisted particle swarm optimization (SPSO) strategy is developed.
Abstract
The severe stochastic nature, multi-scale variations, and non-stationary characteristics typical of wind power temporal data creates substantial difficulties for reliable forecasting. To tackle these specific obstacles, this paper develops an advanced hybrid framework for wind power forecasting, VMD-PSO-LPMS-Transformer-ARFIMA, with global hyperparameter optimization based on SPSO. Firstly, the core hyperparameters of the variational mode decomposition (VMD) algorithm are adaptively optimized using particle swarm optimization (PSO), facilitating stable signal decomposition and robust noise suppression. Subsequently, a linear predictability median split (LPMS) method is proposed. This strategy quantifies the predictability of each decomposed component through the forecasting error of a linear proxy model on the validation set and uses this measure to partition the components into high-frequency and low-frequency groups. In the forecasting stage, the low-frequency components characterized by long memory trends are modeled using autoregressive fractionally integrated moving average (ARFIMA), whereas the high-frequency nonlinear fluctuations are captured using a Transformer architecture to model global dependencies. To alleviate the substantial computational expenses frequently encountered during hyperparameter optimization for deep learning frameworks, this study incorporates a surrogate assisted particle swarm optimization (SPSO) strategy. This approach relies fundamentally on a Gaussian Process surrogate model to efficiently streamline the tuning procedure. Experimental validation is performed utilizing wind power datasets collected from a wind farm located in Xinjiang, China, with April and October selected as representative seasons for comparative and ablation analysis. Based on the empirical evaluations, superior stability across varying seasons alongside the highest prediction precision is successfully realized through the developed hybrid architecture.