Smart solar power prediction using an FFT-Infused ShuffleNet regressor: A high-accuracy lightweight framework

• Introduces “FFT-ShuffleNet,” a novel hybrid model combining FFT preprocessing with a lightweight ShuffleNet for solar power prediction. • FFT preprocessing boosts model performance, cutting MSE by 65% and lifting R 2 from 0.87 to over 0.97 in PV output prediction. • Data shuffling improves model generalization and robustness, leading to MAPE gains of up to 50% over unshuffled models. • Combining FFT and data shuffling yields significant gains in prediction accuracy, robustness, and flexibility across diverse zones. • The FFT-ShuffleNet model is highly accurate, compact, and computationally efficient, ideal for real-world energy management systems. Accurate photovoltaic (PV) power prediction is essential for grid stability, yet state-of-the-art deep learning models present a persistent trade-off: high accuracy often comes with significant computational costs, limiting real-world deployment. This paper introduces the FFT-ShuffleNet, a novel framework designed to resolve this challenge by achieving high accuracy within an exceptionally lightweight architecture. Our approach creates a powerful synergy by using the Fast Fourier Transform (FFT) to extract critical frequency-domain features, which are then processed by a ShuffleNet-based regressor adapted for time-series analysis. The proposed model was first validated on Global Horizontal Irradiance (GHI) forecasting and subsequently applied to PV power output prediction across multiple climatic zones in China. Results consistently show that FFT significantly improves model performance, reducing Mean Squared Error (MSE) by up to 65% and boosting the coefficient of determination (R 2 ) from 0.87 to over 0.97 in several regions. Data shuffling further enhances generalization, with Mean Absolute Percentage Error (MAPE) gains of up to 50% compared to unshuffled models. This work not only advances the state of the art in PV forecasting but also delivers a solution that uniquely balances high accuracy with the low-latency, low-memory footprint required for practical deployment in real-world energy management systems. Future work will focus on extending this high-efficiency strategy to larger deep learning networks.