Digital-twin versus statistical PV modeling: The role of meteorological uncertainty

PV plant metadata has steadily improved in Germany, including module coordinates, orientations, and inverter limits, as well as data on self-consumption and storage. With this information, it is possible to make increasingly granular predictions of regional PV and model more and more plant characteristics deterministically instead of statistically. In the long term, the goal of Redispatch 2.0 in Germany envisions a fully dynamic power management of millions of individual plants to minimize bottlenecks in a landscape of renewables. Such digital twinning however makes certain assumptions about the accuracy and resolution of meteorological observations and forecasts. Assuming a plant has a specific angle at a specific location, for example, makes it more sensitive to the solar position and meteorological uncertainty than when its capacity is dispersed over a distribution of possible angles or over an area more representative of the meteorological variability.

This study explores this contradiction, that more deterministically detailed metadata can lead to less deterministically accurate PV estimates. Two regional PV models with the same underlying physics are compared, one a postal-code-based probabilistic model and another a full digital twin of Germany. Despite the plant-specific details in the digital twin, it performs worse than the statistical model against German meter data. We investigate the differences by modifying the available metadata and plant dispersion in the systems. Moreover, the experiments probe the role of meteorological variability by comparing both numerical weather predictions and satellite data and by artificially reducing the resolution of satellite observations. Lastly, the forecasting systems estimate not only PV production but also model self-consumption and storage reductions to grid feed-in. These very nonlinear aspects are particularly interesting in the context of averaging the local variability.

The results suggest that the benefits of digital twinning can first be realized if the meteorological resolution and uncertainty can match the specificity of the assumed plant characteristics. Otherwise, the PV installations can only be treated deterministically if the meteorological data is probabilistic, which is computationally expensive for digital twins. This probabilistic treatment is feasible with the physics-based model presented here, but poses intense technical and computational challenges to future, data-driven digital twins and redispatch.