| Summary: | Accurate power prediction for floating photovoltaic (FPV) systems is essential for reliable grid integration, yet existing forecasting methods suffer from suboptimal parameter configuration, excessive computational demands, and poor performance under dynamic conditions. This study presents a novel Teaching–Learning-Based Optimization enhanced Extreme Learning Machine (TLBO-ELM) framework that achieves optimal parameter configuration without algorithmic tuning while maintaining computational efficiency for real-time deployment. Unlike conventional approaches, the TLBO-ELM integrates the parameter-free optimization capabilities of TLBO with ELM’s analytical weight calculation, systematically optimizing input weights and biases through simulated teaching–learning processes, thereby eliminates random initialization sensitivity while ensuring global optimization convergence. The framework was validated using comprehensive year-long operational data from a Malaysian FPV installation and benchmarked against established optimization-based ELM variants (GA-ELM, PSO-ELM, BMO-ELM). Results demonstrate superior performance with R2 = 0.9384, RMSE = 7.82 kW, and MAE = 3.17 kW significantly outperforming all benchmark methods with statistical significance at p < 0.001. Notably, comprehensive interpretability analysis revealed that three-phase electrical current measurements dominate predictive capability, enabling cost-effective monitoring strategies for resource-constrained marine environments. The framework maintains robust performance across diverse operational conditions while acknowledging specific limitations during extreme weather transitions and rapid power fluctuations. This work represents the first parameter-free optimization approach for FPV forecasting that simultaneously achieves superior accuracy and computational efficiency, enabling practical real-time deployment in resource-constrained marine environments and advancing optimization-driven machine learning for renewable energy applications.
|
|---|