Multi-interval hybrid AI framework for day-ahead solar pv forecasting with accuracy-efficiency benchmarking in tropical utility-scale systems

The stochastic PV generation under rapid tropical meteorological variability challenges grid stability and planning. Unlike conventional studies that typically assess individual hybrid architectures or a single temporal resolution, this study establishes a unified cross-framework, multi-resolution benchmarking approach that enables hybrid architectures to be evaluated under a common experimental setting. Operational data from a utility-scale Malaysian PV plant are used for day-ahead forecasting. Meteorological correlation analysis at 5-minute, 30-minute, and 1-hour resolutions establishes representative input configurations. Four staged architectures, i.e., ANN-SVR-FNN, CNN-SVR-FNN, LSTM-SVR-FNN, and CNN-LSTM-FNN, combine feature representation, prediction learning, and fuzzy-inference refinement. Genetic Algorithm-Particle Swarm Optimization (GA-PSO) and its Bayesian-enhanced variant (GA-PSO-BO) support prediction-stage hyperparameter tuning. Predictive performance is assessed using R², MAE, and RMSE alongside peak memory, model-development time, inference latency, and estimated multiply-accumulate operations. ANN-SVR-FNN optimized with GA-PSO at the 1-hour resolution using all irradiances and ambient temperature achieves the highest R² of 0.996204 and lowest RMSE of 964.81 W. CNN-LSTM-FNN optimized with GA-PSO–BO at the 30-minute resolution using plant irradiance and ambient temperature achieves the lowest MAE of 555.90 W. Stronger performance at coarser resolutions is consistent with temporal smoothing, while competitive reduced-input configurations highlight the importance of feature relevance rather than input dimensionality or increasing architectural complexity alone. Across the shortlisted configurations, measured inference latency ranges from 0.242 to 1.164 ms per forecast, suggesting that model inference is unlikely to be a computational bottleneck for day-ahead forecasting on the evaluated platform. The findings support metric-specific model selection through joint assessment of predictive accuracy and computational requirements under variable tropical conditions.