Transferability of data-driven power estimate models from land to floating PV

Accurate power generation estimation is essential for the bankability and design optimization of solar photovoltaic (PV) systems, including emerging technologies such as floating PV (FPV)—but remains challenging. Research to date has largely focused on short-term forecasting. Long-term estimation, however, continues to depend on models like SAM and PVsyst, which systematically underestimate FPV yield by approximately 9 % because of coarse environmental representations. In contrast, estimation error for land-based PV (LPV) systems using these tools is typically limited to 1–5 %. To address this gap, we evaluate two machine learning models—XGBoost and Long Short-Term Memory (LSTM)—for high-resolution FPV power estimation and use XGBoost to assess transferability from LPV systems. We employ a 20-month tropical dataset with minute-level resolution, capturing environmental and power output data from co-located FPV and LPV systems. This fine granularity enables accurate modelling of rapid environmental dynamics. Both machine learning models outperform the pvlib baseline in FPV self-estimation, achieving a mean absolute error of approximately 30 W, a root mean square error of 83 W, and a coefficient of determination greater than 99 % for a 5.6 kW system. SHAP (SHapley Additive exPlanations) analysis highlights irradiance, temperature, and time as key influencing features. In the transferability analysis, an XGBoost model trained on only one month of LPV data achieved R 2 = 99.2 % when estimating FPV output over the entire 20-month period. However, differences in surface albedo (rooftop vs. water) and longer LPV training durations each reduced transferability accuracy. These findings demonstrate the potential of data-driven approaches for reliable FPV performance estimation and introduce a scalable LPV-to-FPV transferability framework for emerging PV applications. • Used a 20month, minute-level FPV and LPV dataset for high-resolution modelling. • ML models outperformed pvlib with R²> 99 %, MAE ≈ 30 W, RMSE ≈ 83 W on FPV data. • LPV-trained model achieved 99.2 % R² on FPV using just one month of training data. • Different albedo behaviour and long LPV training reduced FPV transferability.