Q-DASTNet: Integrating quantum optimization and deep learning for robust solar energy forecasting in smart grid systems

Document Type

Article

Publication Title

Results in Engineering

Abstract

The proper working of the smart grids depends on the ability to predict the levels of solar energy since everything becomes balanced in terms of electricity distribution, energy usage is optimized, and renewable resources may be integrated with minimum inconveniences. Nevertheless, traditional machine learning and heuristics learning models tend to fail in accounting the far more volatile and non-linear behavior of solar irradiance especially in dynamic variable environmental conditions. To address these shortcomings, this paper presents a new Q-DASTNet, which is a novel Quantum-driven Deep Attention-based Spatiotemporal Network focused on solar energy forecasting. The main novelty of Q-DASTNet is its unusual architecture of combining three components into one hybrid model: (i) Convolutional layers able to extract meaningful spatial features out of databases on meteorological conditions and solar parameters; (ii) Gated recurrent memory structures, e.g., GRUs or LSTMs, enabling the visualization of long-term dependencies; and (iii) Self-attention mechanisms designed to prioritize the meaning of time intervals and environmental variables. This deep architecture is also optimized with the help of quantum-inspired optimization algorithm that adds a probabilistic search mechanism founded on quantum concepts that speed up the convergence, escape local minima as well as accelerate generalization of the model. The combination of quantum optimization and attention-guided deep learning is a radical break away with conventional methods. The model was assessed thoroughly with Kaggle solar energy data set enhanced with Fingrid energy data. It attained the minimum Mean Absolute Error (MAE) of 0.0098, the maximum Coefficient of Determination (R²) of 0.9864, and the minimum Root Mean Squared Error (RMSE) of 0.0128, outclassing all baseline models.

DOI

10.1016/j.rineng.2025.107306

Publication Date

12-1-2025

This document is currently not available here.

Share

COinS