The impact of temporal smoothing on the accuracy of separation models

Abstract

Passing clouds and the phenomenon of cloud enhancement (CE) can cause severe fluctuations in the incoming solar radiation. The modeling of PV systems’ operation is conditioned by the ability of forecasting and estimation models to capture such phenomena. In time-averaged datasets, these high-frequency events can be completely smeared out. Thus, it is important to study the impact of temporal smoothing on the accuracy of solar irradiance models. This paper focuses on empirical separation models.In the first part, the frequency of CE episodes in the measured data is analyzed, finding an exponential decrease with increasing data timescale. In the second part, the impact of temporal smoothing on the performance of separation models developed on 1-h or 1-min data is assessed. The results show that the aggregated nRMSE of the hourly models decreases from 37.7% to 28.6% between smallest and largest considered timescale. Under the same scenario, the nRMSE of the minute models decreases from 32.1% to 29.7%. Additionally, the hourly models tend to underestimate the measured diffuse fraction, the aggregated nMBE varying between –11.3% and –1.1%, while the minute models tend to overestimate the measured data, the aggregated nMBE ranging between 2.8% and 8.4%, as the averaging interval is increased.The performance of the models is found to have a strong dependence on climate type. No dominating model could be found. The models which perform most consistently are: (a) Yang2 and BRL-min in temperate climates, (b) BRL and PB in tropical climates, and (c) PB and BRL-min in arid climates.