Improving the accuracy of solar irradiance forecasts based on Numerical Weather Prediction Bibek Joshi, Alistair Bruce Sproul, Jessie Kai Copper, Merlinde Kay
Why solar power forecasting? Electricity grid management Solar power plant operation Residential energy management P2P trading
Spatial resolution/ horizon (km) PV power and solar irradiance forecasting methods 1000 100 10 1 0.1 Sky Imagery Satellite Imagery Best forecasts from NWP Solar irradiance forecasts Temperature ad wind forecasts PV system model 0.01 PV generation forecasts 0.001 0.01 0.1 1 10 100 1000 Time horizon (hr) Figure: Classification of forecast model based on spatial and temporal resolution (Inspired by Diagne et al. 2013)
Numerical Weather Prediction (NWP) Data products Surface level forecasts Isobaric levels forecasts Source: http://www.jma.go.jp/jma/jma-eng/jma-center/nwp/grid.jpg
NWP errors Data collection and assimilation Grid resolution Cloud representation Error propagation Machine learning models can learn significant portion of the complex atmospheric process
Data sources Attributes NWP Domain Spatial resolution Time steps Description ACCESS-R APS2 Australian continent and surrounding seas 0.11 (~12 km) Hourly out to 72 hours Data period: Aug 2017 Jul 2018 Location of ground measurement stations
NWP Errors x x Obs: Observations C: ACCESS-C NWP R: ACCESS-R NWP CS: Clear sky model
Input dataset Surface level forecasts Global horizontal irradiance Direct horizontal irradiance Diffuse horizontal irradiance Air temperature Total cloud cover High cloud cover Middle cloud cover Low cloud cover Specific humidity Mean sea level pressure Wind speed (u component) Wind speed (v component) Heights of cloud bases for different oktas Isobaric levels forecasts Air temperature Relative humidity Wind speed (u component) Wind speed (v component) Vertical wind speed Geopotential height Isobaric levels (hpa): 500,600,700,800,850,900,950, 1000 Solar position variables Solar zenith angle Extraterrestrial irradiance Measured Global Horizontal Irradiance (GHI) Hourly aggregated GHI from ground stations
Inputs Relevant surface level forecasts Relevant isobaric levels forecasts Solar position variables Total number of features = 63 Improving GHI forecast accuracy: Machine learning framework Feature selection Mutual Info Regression Principle Component Analysis Gradient Boosting Regression (GBR) Optimization of model parameters Training, validation and testing Hourly average GHI forecasts (1-24h) Immediate past 3 months data used for training Training stage Target hourly GHI The model is trained separately for each site on a daily rolling basis to forecast 24-h window.
Dynamic selection of hyperparameters of GBR Site: Adelaide Site: Adelaide Trained using default hyperparameters Computation time per site: ~ 220 seconds Trained using dynamic selection of hyperparameters Computation time per site: ~415 seconds
Forecast vs Observed GHI: All sites R 2 =0.84 R 2 =0.88 R 2 =0.91 Raw model Model 1 Inputs to machine learning Surface level forecasts Solar position variables Model 2 Inputs to machine learning Surface level forecasts Isobaric level forecasts Solar position variables
Performance metrics for GHI forecasts: All sites 40% 30% 20% 10% 0% -10% NMBD NMAD NRMSD Persistence Raw Model1 Model2 NRMSD for Model 1 and Model 2 reduced by 17% and 31%, respectively, with respect to the raw model NMBD: Normalized Mean Bias Deviation NMAD: Normalized Mean Absolute Deviation NRMSD: Normalized Root Mean Squared Deviation
Performance metrics: individual sites MBD (W/m 2 ) 20 10 0-10 -20-30 -40-50 -60-70 Adelaide Alice Springs Broome Cape Grim Darwin Geraldton Kalgoorlie Boulder Learmonth Melbourne Rockhampton Townsville Wagga Wagga Raw Model 1 Model 2 Persistence RMSD (W/m 2 ) 250 200 150 100 50 0 Adelaide Alice Springs Broome Cape Grim Darwin Geraldton Kalgoorlie Boulder Learmonth Melbourne Rockhampton Townsville Wagga Wagga Raw Model 1 Model 2 Persistence
Forecast performance in different sky conditions Overcast Cloudy Clear kk tt = GGGGII tt GGGGII cccc wwwwwwwww GGGGII tt iiii oooooooooooooooo GGGGGG aaaaaa GGGGII cccc iiii estimated cccccccccc ssssss GGGGGG Observed
Sample time series Site: Adelaide
Summary and conclusions Information from forecast variables other than solar irradiance can be leveraged to improve solar irradiance forecast accuracy. Isobaric forecasts contribute to additional improvement in forecast accuracy relative to the models using surface level forecasts only. Boosting methods can easily overfit the training data. Dynamic selection of hyperparameters can mitigate overfitting issues.
Thank you! Acknowledgements This research is funded by Cooperative Research Centres for Low Carbon Living Ltd (CRC-LCL), an Australian Government Initiative and supported by Solar Analytics Pty Ltd, an automated fault monitoring service for rooftop solar energy systems The ACCESS NWP forecasts were provided by the Australian Bureau of Meteorology for this research.