MACHINE LEARNING APPLICATIONS IN RENEWABLE ENERGY: PREDICTIVE ANALYTICS FOR SOLAR CELL PERFORMANCE OPTIMIZATION AND ENERGY YIELD FORECASTING

Abstract

This systematic review synthesizes contemporary evidence on machine-learning (ML) applications for solar photovoltaic (PV) performance optimization and energy-yield forecasting, spanning algorithms, data infrastructures, evaluation practice, and operational integration. Following PRISMA guidelines, we screened multidisciplinary databases and included 214 empirical studies for qualitative synthesis. Findings reveal a consistent accuracy hierarchy: tuned tree-based ensembles are the most dependable and computationally economical for day-ahead, tabular mappings of numerical weather prediction and plant telemetry; deep neural architectures (e.g., CNN/LSTM and hybrids) dominate minute-to-hour nowcasting when inputs are image- or sequence-rich; and physics–ML hybrids improve robustness and physical plausibility under regime shifts or sparse data. Cross-regional validation exposes systematic optimism in single-site splits; region-out testing typically increases error, while transfer learning and domain adaptation halve that penalty in many cases. Data quality emerges as the performance ceiling: standards-aligned sensing, explicit soiling treatment, synchronized timestamps, and streaming feature engineering yield error reductions comparable to algorithmic gains. IoT and big-data stacks—edge inference for sub-minute latency paired with cloud-based training, monitoring, and drift management—prove critical for real-time operation. Beyond forecasting, image- and I–V–based diagnostics achieve high scores for fault detection, and sequence-aware prognostics support remaining-useful-life estimation. Explainability layers (e.g., attribution or saliency) facilitate adoption without sacrificing accuracy, especially when coupled with physics-guided features and probabilistic outputs for grid dispatch and storage control. Overall, durable value arises from aligning horizon-appropriate models with disciplined data pipelines, climate-aware evaluation, and production-grade MLOps; future progress hinges on broader geographic coverage, open benchmarks, advances in transfer/physics-informed learning, and governance that ensures transparency, security, and market interoperability.