A Multi-Phase Deep Learning Framework for Multi-Step Short-Term Wind Power Forecasting in Presence of Uncertainties

Abstract

With the expanding share of wind energy in power grids, accurate forecasting has become critical for maintaining system stability and operational efficiency. Notwithstanding, forecasting accuracy is compromised by uncertainties from fluctuating wind speeds and meteorological conditions. This paper proposes a novel multi-phase short-term wind power forecasting framework (multi-step ahead forecasting over a 1-hour horizon). Thus, decomposition of the wind power signal and feature extraction are initially implemented using Variational Mode Decomposition (VMD) and Principal Components Analysis (PCA), respectively, aiming to enhance input quality and reduce computational burden. The proposed forecasting model is built on a hybrid DL architecture merging a Convolutional Neural Network (CNN), Attention Mechanism (AM), and Deep Feedforward Neural Network (DFFNN). Given the impact of decomposition levels and extracted PCA components count on forecasting performance, a search-based scheme is developed to explore a pre-defined space (maximum decomposition level and extracted components count) to determine the optimal configuration for each interval. In the next phase, a Fuzzy Decision-Making (FDM) technique is employed to select a balanced and optimal configuration for the proposed model across the year. To demonstrate the proposed architecture’s efficacy and generalizability, the model is tested on two real-world data from La Haute Borne wind farm in France and Hill of Towie wind farm in Scotland. Results demonstrate that the proposed architecture with the selected configurations achieves significant accuracy and generalization, with average NRMSEs and NMAEs values of 0.428% and 0.333% for La Haute Borne wind farm and 0.502% and 0.381% for Hill of Towie wind farm.