Hybrid Intra-hour Solar Power Forecasting using Statistical and Skycam-based Methods
Jing Huang1, Maryam Khan2, Yi Qin1, Sam West2
1CSIRO Oceans and Atmosphere, Canberra, Australia
/2CSIRO Energy Technology, Newcastle, Australia

The importance and urgency of solar power forecasting grow with the current trend of increasing solar power installations all over the world.  Accurate solar power forecasts facilitate grid operators to reduce the running cost of an electric grid with a high penetration of variable solar power generation and to enhance the quality of service for electricity consumers. In particular, Australia is currently striving for developing more accurate 5-minute ahead power forecasting for wind and solar farms operating in its National Electricity Market (NEM), which provides direct motivation for this work.

Both statistical methods and skycam-based approaches are capable of forecasting solar PV power generation at minutes ahead. Statistical methods are self-adaptive but blind due to lack of physical information. Skycam-based approaches are physical and thus susceptible to many factors such as the quality of cloud motion vector algorithm, the quality of power conversion model and characteristics of local clouds. We propose and test a hybrid solar PV power forecasting model, termed as Conditional ARX (CARX) model, which optimally combines statistical and skycam-based forecasts using high-frequency measurements in Canberra, Australia.

We show CARX’s capability to produce accurate forecasts seamlessly from 10-s to 10-m ahead.  The hybrid model relies on an empirical clear-sky model for solar power and the identification of three condition variables (i.e. clear-sky index, temporal gradient of clear-sky index and solar zenith angle), which are able to separate and model characteristic events associated with them. It significantly overperforms both its statistical component and its skycam component alone, achieving a relative RMSE reduction (forecast skill) of 19% against smart persistence (i.e. persistence of clear-sky index) at 5-m ahead. The relative importance of skycam-based forecasts in CARX and CARX’s overall forecast skill generally increase with forecast horizon, confirming the value of the cloud-resolving capability of skycam-based approaches in the long run.