Answers are grounded in this paper and cite their sources.
Proper photovoltaic (PV) power forecasting supports grid stability, efficient dispatch, and economic performance. While most forecasting models use rigid time smoothing, which eliminates multi-scale volatility, in this study, we propose a High–Mid–Low Triangular Echo State Network (HMLT-ESN) optimized by a Quasi-Oppositional Learning Cyclonic Foraging Beaver Behavior Optimizer (QCBBO). HMLT-ESN partitions photovoltaic sequences into high-, mid-, and low-frequency components using Variational Mode Decomposition (VMD). Each component is processed by a sub-reservoir, and each sub-reservoir uses different activation functions: TreX for transients, tanh for periodicities, and sigmoid for trends. Rather than using a fixed leakage rate, we introduce an attention-driven diagonal matrix that adjusts each timescale blockwise according to its instantaneous activation energy. Recurrent topology is restricted to an equilateral triangular lattice that provides an analytically tractable spectral radius. The circle map achieves better coverage over the normal input domain and feature discrimination, while the QCBBO runs autonomously to avoid local optima. Tested on a 20 MWp plant in northern China, the model predicts 15 min power at 15 min resolution using 90 min windows. The HMLT-ESN reduces Mean Absolute Error (MAE) by 41.57% and Normalized Mean Square Error (NMSE) by 51.73% compared to ESN, acting as a robust dispatch tool for renewable-heavy power systems.
No takes yet. Share an insight, caveat, or question.
Yao et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: