A comprehensive review of deep learning for solar nowcasting: Enhancing accuracy, reliability, and interpretability

Solar nowcasting (0–6 hour horizons) is critical for grid stability, ramp-rate control, and energy storage optimization as renewable penetration accelerates. This systematic review analyzes 120 peer-reviewed studies (2015–2024) selected via PRISMA protocol from 8245 initial records, evaluating CNN-based spatial feature extraction, RNN-based temporal modeling, and hybrid multi-modal fusion architectures alongside emerging paradigms including federated learning, diffusion models, physics-informed neural networks, and foundation models. Modern deep learning achieves 20–40% improvement over persistence baselines (skill scores 0.25–0.45), with physics-aware designs substantially reducing violations of radiative constraints. However, critical gaps persist: only 18% of studies release code publicly, preprocessing pipelines remain undocumented, probabilistic evaluation using proper scoring rules (CRPS, Brier score) with calibration diagnostics is inconsistently applied, and benchmark datasets concentrate in North America and Europe, limiting generalizability. We establish explicit connections between forecast skill and operational value (reserve costs, curtailment reduction, ramp compliance, battery cycling) and quantify deployment constraints (inference latency < 60s, energy consumption 5–300W, edge versus cloud architectures). Key recommendations include mandatory release of preprocessing pipelines with datasets, standardized probabilistic evaluation protocols with condition-specific analyses, physics-informed architectures with radiative constraints, multi-objective optimization balancing accuracy against computational cost and carbon footprint, and federated learning for privacy-preserving collaboration. This review provides an evidence-based roadmap toward reproducible, physically consistent, and operationally valuable solar nowcasting essential for reliable renewable energy integration.