This review combines bibliometric mapping with a critical narrative discussion of machine learning in solar-energy systems. Scopus provided two related, separately analysed bibliographic sets. The photovoltaic search used “Machine Learning AND Photovoltaics” in the article title, abstract and keywords fields. Its export contains 1,954 distinct Scopus EIDs from 2010–2026: 1,143 articles, 606 conference papers, 146 reviews and 59 other records. Two records are marked “Retracted” and are retained in that descriptive total but not used as technical evidence. A broader search for “Machine Learning AND Solar Energy” returned 6,767 results. Because the two searches cover different, potentially overlapping populations, we interpret their VOSviewer maps separately and do not combine their counts. The 2026 entries represent a partial publication year and are not treated as a complete annual trend. The technical synthesis draws on relevant published studies of irradiance and photovoltaic forecasting, fault detection, energy management and grid applications; it is not a screened subset of either bibliometric set. We interpret reported results by target variable, forecast horizon, input source, climate, sample size, chronological validation, prevention of preprocessing leakage, benchmark choice, uncertainty assessment and testing at another site or period. These criteria guide the discussion rather than constitute a numerical quality score or a formal risk-of-bias assessment. The retained materials do not establish the exact search dates, full executed queries, filters, duplicate-removal log or screening decisions. VOSviewer version 1.6.20 is visible in the analysis records, but the original counting and normalisation settings, thesaurus and native network files are unavailable for verification. We therefore present descriptive maps and a selective technical synthesis, without a PRISMA flow or claims of systematic-review completeness. These constraints limit independent reproduction and cross-study comparisons.

Figure 3 displays the original country and keyword visualisations for the two separate searches. The keyword maps show solar forecasting and machine-learning terms alongside PV, grid-management and meteorological terms. These are qualitative readings of the displayed maps, not evidence of model superiority, thematic evolution or operational adoption. The earlier analysis reported 16,018 distinct keywords, 1,626 with at least five occurrences and selection of 1,000 by total link strength. The underlying term list and VOSviewer processing log were unavailable, so these intermediate counts and threshold choices could not be independently verified from the photovoltaic csv. The photovoltaic country-selection screenshot lists the five countries with greatest displayed total link strength (TLS): China (475 documents; TLS 245), Saudi Arabia (123; 227), United States (289; 194), United Kingdom (104; 130) and India (278; 121). Their TLS values total 917, giving the percentages in the lower pie of Figure 3 after rounding: 26.7%, 24.8%, 21.2%, 14.2% and 13.2%, respectively. These percentages describe the combined TLS of only those five selected countries, not shares of publications. The upper panel represents the separate solar-energy search and is not pooled with the photovoltaic data. A co-authorship link reflects collaboration within a retrieved set; it does not establish research quality or model performance.

Solar-resource information supports decisions from long-term site selection and capacity planning to short-term dispatch, storage scheduling, and grid balancing. Figure 4 presents a visual overview of the solar energy resources and management framework adopted in this study. ML is particularly useful because it can integrate ground observations, satellite products, NWP outputs, remote-sensing indices, and historical operating data while representing nonlinear atmospheric variability (Sulaiman and Mustaffa 2024; Sun et al. 2022; Liu et al. 2022). For practical solar-energy modelling, the principal surface components are GHI, DNI, and DHI, whose physical meaning and measurement constraints determine the appropriate target, input data, and downstream application (Liu et al. 2022).

At the surface, solar radiation is modified by absorption, scattering, reflection, and solar geometry. The three principal components satisfy the horizontal-plane relationship , where theta z is the solar zenith angle. This identity provides a useful physical-consistency check for measurements and component-wise ML predictions (Liu et al. 2022).

Solar-energy management is constrained by a coupled atmosphere-measurement-technology problem. Rapid weather transitions and cloud motion create short-term ramps that historical time series or coarse NWP fields may miss (Juncklaus Martins et al. 2022). Ground sensors introduce calibration drift, soiling, shading, and data gaps, with DNI and DHI particularly affected by sparse measurement coverage (Daniel-Durandt and Rix 2023; Esposito et al. 2024). Models also face geographical shift because irradiance statistics differ across maritime, arid, monsoon, mountain, and urban environments (Kumar and Suzuki 2019). Finally, the irradiance-to-power relationship is nonlinear because of cell temperature, spectral response, partial shading, soiling, ageing, mismatch, and inverter controls (Rahman et al. 2023). These interacting sources of uncertainty limit model transferability and motivate quality control, domain adaptation, uncertainty-aware learning, and physically consistent prediction (Liu et al. 2022; Dai et al. 2025; Shata et al. 2025).

SE management spans forecasting, control, diagnostics, optimisation, demand coordination, and grid support (Singh et al. 2024). These tasks share data but differ in time scale and decision objective. Forecasting estimates irradiance or PV power from seconds to days ahead; MPPT seeks the instantaneous operating point that maximizes module output; fault diagnosis detects soiling, hotspots, degradation, inverter failures, and wiring anomalies; farm optimisation coordinates trackers, inverters, maintenance, curtailment, storage, and dispatch; demand response aligns flexible loads with variable PV supply; and grid-integration tools support ramp-rate control, voltage regulation, reserve planning, and stability assessment. ML contributes by learning nonlinear relationships, extracting patterns from images and SCADA streams, and converting forecasts into adaptive control or maintenance decisions. The main value therefore lies not in prediction alone, but in linking prediction to operational action.

Classical methods remain strong baselines for structured meteorological and operational datasets. ANN provide flexible nonlinear approximation; SVM/SVR are effective for small-to-medium datasets with carefully tuned kernels; decision trees are interpretable but unstable as single learners; RF improve robustness through bagging and feature randomization; GPR provides predictive distributions and explicit uncertainty; and fuzzy systems or ANFIS incorporate rule-based reasoning. Their lower data and computational requirements often make them attractive for site-specific applications and embedded systems (Mohammed et al. 2025).

Ensemble learning combines multiple predictors to reduce bias, variance, or model-specific error. Bagging and RF stabilize high-variance learners; boosting methods such as XGBoost, LightGBM, and CatBoost sequentially correct residual error; and stacking uses a meta-learner to combine heterogeneous models. For structured solar datasets, ensembles often provide an effective accuracy-computation compromise and can outperform single ANN or SVR models when feature quality is high (Ramírez-Rivera and Guerrero-Rodríguez 2024; Anonto et al. 2025).

DL is most advantageous when large, high-resolution, sequential, spatial, or image-based datasets are available. LSTM networks model temporal dependence in irradiance and power series; CNN extract spatial features from sky or satellite imagery; CNN-LSTM hybrids integrate spatial and temporal information; and attention-based transformers can model long sequences and multi-step dependencies. Their benefits, however, must be weighed against training cost, data requirements, and reduced interpretability (Barancsuk et al. 2025; Seman et al. 2026).

Hybrid models improve learning by combining optimisation, decomposition, or physical knowledge with ML. Metaheuristics such as GA and PSO tune network weights, kernels, hyperparameters, or feature subsets; wavelet, STL, and EMD decompositions separate multiscale or seasonal structure before prediction; and statistical-ML hybrids combine complementary assumptions. Physics-informed ML is especially promising because radiative, electrical, or energy-balance relations can be embedded as constraints or regularisation, improving consistency and potentially reducing extrapolation error (Santos et al. 2024). Figure 6 delineates the structural frameworks of the PSO and GA algorithms.

Emerging methods shift the field from isolated forecasting toward distributed and adaptive energy intelligence. RL learns control policies for PV-battery dispatch, demand response, and MPPT; GNN represent spatial and electrical relationships among PV strings, inverters, feeders, or microgrids; federated learning may enable collaborative training across sites under stated privacy and communication assumptions; and digital twins may couple live measurements with virtual system models for monitoring, prediction, and what-if analysis. The reviewed evidence offers application-specific examples for reinforcement learning (Fang et al. 2025) and urban graph modelling (Li et al. 2025); broad deployment benefits of federated or agentic systems remain unestablished here. Safety, communication, cyber-security, validation and computation should be assessed separately. The emerging trends in SE applications using GNN and DT technologies are presented in Table 3.

No single model is universally best. Classical learners remain competitive for moderate structured datasets, ensemble, deep and physics-guided methods have different data and validation requirements. Their relative performance cannot be ranked across heterogeneous studies. Model comparison should therefore consider accuracy, uncertainty, robustness, transferability, interpretability, and deployment cost together. Figure 7 illustrates key emerging trends in SERM through Agentic AI in panel (a) and Digital Twin technology in panel (b).

SR forecasting predicts GHI, DNI, or DHI across horizons from minutes to days. Very short-term methods benefit from sky cameras, satellite imagery, cloud motion, and recent measurements; intraday forecasts often combine temporal ML with meteorological inputs; and day-ahead forecasts increasingly post-process NWP outputs. LSTM, CNN, ensembles, decomposition hybrids, and metaheuristic-tuned models can improve accuracy, but fair comparison requires climate-specific baselines, persistence benchmarks, and consistent forecast horizons. The strongest studies treat forecasting as an uncertainty problem rather than reporting point-error metrics alone.

PV power prediction adds module temperature, system design, degradation, inverter behaviour, shading, and soiling to irradiance uncertainty. Direct power models can learn the full weather-to-power mapping, while staged models first forecast irradiance and then convert it to power. RF and ensembles are often effective for structured plant data, whereas LSTM and hybrid models are useful when temporal dependence is strong. Operationally useful models should remain reliable during ramps, rare weather, curtailment, and equipment changes rather than only under average conditions.

Solar-farm management extends prediction into action. ML can identify abnormal energy loss, prioritize cleaning and maintenance, forecast component failure, optimize trackers and inverter loading, coordinate storage, and support curtailment or dispatch decisions. RL and digital-twin approaches can further integrate forecasting and control, but deployment requires explicit constraints, safe fallback behaviour, and evaluation against economic and reliability outcomes rather than prediction error alone.

Solar resource assessment increasingly combines ground observations with satellite, reanalysis, GIS, and climate-projection data. ML can correct bias, fill spatial gaps, downscale coarse products, and rank candidate sites. For long-term planning, transferability and climate non-stationarity are more important than small gains on a single historical test set; validation should therefore span multiple locations, seasons, and where possible future climate conditions.

In smart grids and urban energy systems, ML-based load forecasts can be coordinated with PV forecasts, batteries, flexible buildings, and electric-vehicle charging. The resulting problem is inherently multivariate and decision-oriented: forecast errors affect reserve needs, curtailment, comfort, electricity cost, and network constraints. Future systems therefore require joint probabilistic forecasting and control rather than independently optimising solar generation and demand.

Common metrics include MAE and RMSE for absolute error, MAPE or normalized RMSE for scale-independent comparison, R² or correlation for association, and bias measures for systematic over- or under-prediction. No metric is sufficient alone: RMSE emphasises large errors, MAPE can be unstable near zero, and high R² does not guarantee low operational error. For real deployment, studies should also report uncertainty calibration, inference time, model size, robustness to missing or shifted data, and performance during high-impact ramps. Evaluation should use temporally separated testing and, where relevant, external-site validation rather than random splits that leak seasonal information.

Model strengths are application-specific. SVM and GPR are attractive for smaller datasets, with GPR adding uncertainty estimates but scaling poorly to very large samples. RF and boosting methods are robust on structured data and offer useful feature importance, although tree ensembles may struggle to extrapolate beyond the training range. ANN and LSTM capture nonlinear temporal behaviour but require careful tuning and adequate data. CNN and transformers are powerful for imagery and long sequences but are computationally intensive. Hybrid and physics-informed models can improve accuracy and consistency, yet additional complexity is justified only when gains persist under external validation. Interpretability, maintainability, and computational cost should therefore be treated as model-selection criteria, not afterthoughts.

Reliable benchmarking depends on accessible, quality-controlled datasets with clear metadata. NREL and other public solar archives support reproducibility, but many published studies still use private single-site datasets, inconsistent preprocessing, and different train-test periods. Progress would be accelerated by benchmark suites spanning climates, forecast horizons, sensor types, PV technologies, and operational disturbances, together with standard persistence, physical, and ML baselines. Such benchmarks should include missing data, sensor drift, and distribution shift rather than only clean historical records.