Intelligent Spectral Irradiance Forecasting in IoT-Driven Smart Solar Grids: A Hybrid Metaheuristic–Neural Architecture with Fourier- Transformed Matrix Embeddings for Energy Prediction

Abstract

Accurate forecasting of solar spectral irradiance remains a critical challenge in the operational stability and energy
dispatch planning of smart solar grids. Conventional machine learning models often fail to capture the high-frequency
spectral dynamics and non-stationary behavior inherent in solar irradiance data, especially under variable atmospheric
conditions. To address this limitation, we propose a novel hybrid metaheuristic–neural architecture that integrates deep
learning with bio-inspired optimization and spectral signal decomposition. The core innovation lies in the formulation of
Fourier-transformed matrix embeddings (FTMEs), which encode time-series irradiance measurements into structured
spectral–temporal representations. These embeddings serve as input to a deep recurrent neural network (RNN) whose
hyperparameters are dynamically tuned via an enhanced grey wolf optimizer (GWO). Deployed within an Internet of
Things (IoT)-enabled monitoring framework, the proposed system enables real-time, high-resolution irradiance
forecasting across multiple spectral bands. Experimental validation using ground-based spectral irradiance datasets from
the National Renewable Energy Laboratory (NREL) demonstrates a mean absolute percentage error (MAPE) of 2.13%
and a normalized root mean square error (nRMSE) of 0.018 outperforming state-of-the-art benchmarks by 12–19%. The
architecture further exhibits robust generalization across diverse climatic zones, supporting its deployment in nextgeneration
smart solar grids for efficient energy prediction and grid integration.
Keywords: Hybrid metaheuristic–neural algorithms, solar spectral irradiance, IoT-enabled smart grids, deep learning,
Fourier series, matrix transformation, energy forecasting, grey wolf optimizer, recurrent neural networks.