A Novel Combined Objective Hybrid Framework for Ultra-Short-Term Ramp Event Forecasting in Photovoltaic and Wind Power Generation
Abstract
Reliable forecasting of photovoltaic (PV) and wind power generation, particularly in ultra-short-term and short-term forecasting horizons, constitutes an essential tool for grid stability and the effective management of electric power systems with high renewable energy sources penetration. However, machine learning models trained with standard objective functions such as mean squared error minimization tend to produce smooth forecasting curves and thus fail to predict abrupt power fluctuations, i.e., ramp events, which threaten grid stability. In this paper, a hybrid forecasting framework is proposed that integrates the ramp event detection capability directly into the training process. The main predictor is a Long Short-Term Memory network optimized by a new hybrid algorithm combining advanced Simulated Annealing with Particle Swarm Optimization and trained with a novel combined objective function that aligns correct ramp event detection with high prediction accuracy. The proposed framework is applied to ultra-short-term single-step-ahead wind power forecasting and ultra-short-term multi-step-ahead PV power forecasting, utilizing data from a real-world operating wind turbine and PV park, respectively. The experimental results validate the combined objective function’s efficacy in both case studies, as the proposed forecasting framework achieves the highest ramp event prediction capability, while maintaining relatively low average prediction errors compared to several benchmark models.