The complexity and dimensionality of making deterministic photovoltaic power forecasts from ensemble numerical weather prediction

• Leveraging ensemble NWP can improve deterministic PV forecasting accuracy by 5% • Bias correction of PV forecasts is the most important post-processing step. • End-to-end machine learning and post-processed physical model chain have similar accuracy. Ensemble numerical weather prediction (NWP) constitutes a fundamental and reliable way of creating weather forecasts and quantifying their uncertainty. However, converting ensemble solar irradiance forecasts to deterministic photovoltaic (PV) power forecasts is associated with two challenging characteristics, that is, complexity and dimensionality. Complexity is introduced because of the necessary involvement of physical model chains and post-processing tools, both of which require in-depth knowledge of energy meteorology. Dimensionality, on the other hand, arises because one can freely cascade model chains and post-processing tools, each having many alternatives, into 16 distinct conversion workflows, in that, the possibilities multiply. When machine learning is involved, in one way or another, the situation becomes more convoluted. This work provides empirical evidence on the optimal workflow of making deterministic PV power forecasts from ensemble NWP, using four-year data from five utility-scale PV plants in Hungary alongside ensemble NWP forecasts from the European Centre of Medium-Range Weather Forecasts. It is found that (1) using ensemble NWP results in a 5% error reduction over just using deterministic NWP, and (2) bias-correcting the final PV power forecasts is the only indispensable stage of the workflow, which suggests that post-processing irradiance forecasts is not really needed, insofar as the final goal is to forecast PV power.