Metaheuristic-Optimized Deep Learning for Smart Solar Energy Management: A Comprehensive Review

Abstract

The accurate forecasting of solar power generation is a cornerstone of smart energy management, directly affecting dispatch scheduling, battery storage operation, demand response, and the overall stability of grids with high photovoltaic (PV) penetration. Deep learning (DL) architectures, including long short-term memory (LSTM) networks, gated recurrent units (GRU), convolutional neural networks (CNN), and, more recently, attention-based transformers, have become the dominant tool for this task because of their capacity to model the nonlinear and non-stationary behaviour of solar irradiance and PV output. However, the predictive accuracy of these architectures is highly sensitive to their hyperparameters, and manual or grid-search tuning is computationally expensive and rarely converges to a global optimum. Metaheuristic optimization algorithms, inspired by evolutionary, swarm, physics-based, and human-based phenomena, have therefore been widely hybridized with DL models to automate and improve hyperparameter search, feature selection, and, in some cases, network weight optimization. This paper presents a structured and up-to-date review of metaheuristic-optimized deep learning algorithms for solar power and irradiance forecasting within smart energy management systems. We propose a taxonomy of the most commonly used metaheuristics, summarize the architectural characteristics of the deep learning backbones they are paired with, and synthesize quantitative results reported across recent studies, including Fire Hawk Optimization (FHO), the Improved Mountain Gazelle Optimizer (IMGO), the Evolutionary Mating Algorithm (EMA), the Reptile Search Algorithm (RSA), the Grey Wolf Optimizer (GWO), and the Coati Optimization Algorithm (COA). Comparative tables and figures illustrate the relative strengths, computational trade-offs, and reported accuracy gains of these hybrid frameworks, with several studies reporting coefficients of determination (R²) above 0.98 after metaheuristic tuning. The review concludes by identifying open challenges, including computational overhead, limited generalization across climates and PV technologies, and the scarcity of standardized benchmarks, and it outlines promising research directions such as multi-objective and explainable metaheuristic optimization for real-time smart-grid deployment.