Table 1 Comparative summary of existing research on PV power forecasting.

Physical/

mechanistic approaches

NWP-driven irradiance

PV conversion

Physics-grounded; can work with limited historical time-series; transparent assumptions

Sensitive to NWP errors; needs detailed inputs/sensors; calibration overhead

^29,35

Time-series (AR/ARIMA/ARIMAX)

Univariate/multivariate

ARIMA baselines

Simple, fast, strong at short horizons; useful benchmarks

“Inertia” under rapid variability; accuracy drops as horizon grows/sky changes

^13,14,15

Classical ML

(ANN, RF, SVM, GPR, ELM)

Nonlinear learners on engineered features

Handle heterogeneity & nonlinearity; often beat statistical baselines

Tuning/feature engineering; overfitting risk; mixed interpretability

^12,21,22,38

Ensembles

ML + sky-imager + ARIMA blended; ridge-stacked weights

Consistently higher skill than constituents; robust in variable skies

Multi-stream data (imagery + radiometry) and higher implementation complexity

^8,16

Deep learning

(RNN/LSTM/GRU/LSTM)

Sequence models; image-aware CNN hybrids

Strong accuracy with large data; captures temporal–spatial patterns

Data/compute hungry; black-box behavior limits explainability

^23,24,25,28

Hybrid physical +

data-driven

Two-step irradiance forecast + empirical PV model; ML with exogenous physics

Combines complementary strengths; improves generalization

Pipeline complexity; error propagation must be handled carefully

^26,27

Interpretable/

XAI frameworks

Feature selection + SHAP; factor selection + transparent regressors

Identifies influential drivers; auditability; deployment-friendly

Sometimes slightly below best deep ensembles in raw accuracy

^{36, 37}

BIPV application studies

Model tuning, grid search,

K-fold on BIPV strings

Shows performance gains from rigorous tuning/evaluation

Site-/plant-specific constraints; reliance on external met inputs

^33,34,35,36

SVMR vs.

classical regression

Direct comparison: SVMR, Gaussian regression, multivariate regression

SVMR best captured nonlinearities; lowest MAE/MSE; high R2R^2

Kernel/tuning sensitivity; less transparent than linear models

^31,32,39

Reviews/syntheses

Taxonomy & benchmarks (ML, ensembles, DL)

Situates best-in-class methods; highlights ensembles’ short-term gains

Heterogeneous datasets hinder strict apples-to-apples comparisons

³⁵

Interpretable

PV forecasting with

(HFA + Ridge)

HFA to reduce meteorological and

Ridge Regression

Interpretability, Robust to

multicollinearity,

Overfitting resistance, Low cost/

easy deployment

Factor stability assumptions,

No probabilistic output,

Nowcasting under fast cloud dynamics,

Input quality sensitivity

Proposed Method