Exploring super-resolution spatial downscaling of several meteorological variables and potential applications for photovoltaic power

We applied a perfect prognosis approach to downscale four meteorological variables that affect photovoltaic (PV) power output using four machine learning (ML) algorithms. In addition to commonly investigated variables, such as air temperature and precipitation, we also focused on wind speed and surface solar radiation, which are not frequently examined. The downscaling performance of the four variables followed the order of: temperature > surface solar radiation > wind speed > precipitation. Having assessed the dependence of the downscaling accuracy on the scaling factor, we focused on a super-resolution downscaling. We found that the convolutional neural network (CNN) generally outperformed the other linear and non-linear algorithms. The CNN was further able to reproduce extremes. With the rapid transition from coal to renewables, the need to evaluate low solar output conditions at a regional scale is expected to benefit from CNNs. Because weather affects PV power output in multiple ways, and future climate change will modify meteorological conditions, we focused on obtaining exemplary super-resolution application by evaluating future changes in PV power outputs using climate simulations. Our results confirmed the reliability of the CNN method for producing super-resolution climate scenarios and will enable energy planners to anticipate the effects of future weather variability.


FiguresFig
Figures Fig. S1.Time required to train each of four ML-based models over the investigated domain, with a resolution of 0.025° (91 × 92 grid cells).

Fig
Fig. S2 Bias and bias of the 98 th percentile for wind speeds during the test period 2005-2014, for models based on the MLR, RF, ANN, and CNN algorithms (spatial domain-averaged values are provided in Table

Fig
Fig. S7.Comparison of observed and downscaled (CNN-based) precipitation for wet days (precipitation > 1 mm/day) averaged over the test period (2005-2014).Fraction of wet days observed (a) and downscaled (b); precipitation amounts in wet days observed (c) and downscaled (d).

Fig. S8 .
Fig. S8.Solar radiation averaged over the Kanto region during the test period (2005-2014), as in Fig. 4a, except for the difference (CNN output minus observations).

Table S1
Tab. S2.Main hyperparameter settings applied to the four machine learning (ML) algorithms.

Table S3 .
Statistics of the employed ML algorithms, which were trained using 1980-2004 data (except wind speed and surface solar radiation data, see Section 2) and evaluated using 2005-2014 data.The results show the average over the whole study domain during the test period of 2005-2014.Results are expressed as absolute values (i.e., Pr: mm/day, Tas: °C, SW: MJ/m 2 /day, WS: m/s), except for the bias (%) and 98 th percentile bias (%) of Pr.