Towards a More Accurate Machine Learning Multi-Model Ensemble Method for Direct Solar Irradiance Forecasts

- Indicates paper has been withdrawn from meeting
- Indicates an Award Winner
Wednesday, 7 January 2015: 4:30 PM
124B (Phoenix Convention Center - West and North Buildings)
Surya Karthik Mukkavilli, UNSW & CSIRO, Sydney, NSW, Australia; and M. J. Kay, A. A. Prasad, R. Taylor, and A. Troccoli

Handout (3.1 MB)

Concentrating solar power (CSP) plants are projected to be a major contributor in the future clean energy mix based on current operational and planned projects worldwide. Direct solar irradiance is a critical input for CSP scheduling, operations, electricity dispatch and market participation. Ongoing research and development enhancements added to a hybrid Numerical Weather Prediction (NWP) and statistical/artificial intelligence multi-model ensemble being built to produce direct solar irradiance day ahead forecasts for CSP plant operators will be presented. Results from long term study between 2000 and 2012 of Bureau of Meteorology's deseasonalized DNI measurements and aerosol optical depth (AOD) anomalies on the Moderate resolution Imaging Spectroradiometer (NASA Terra satellite) revealed high anti-correlations in north and southeast Australia along with high variability in AOD (~0.03-0.05). However, large biases in existing NWP dynamical models can be further reduced by accurate representation of weather data at mesoscale resolutions, post-processing with machine learning and better representation of clouds schemes, radiative transfer codes, localised aerosol climatologies and chemical transport. In this project a combination of high spatiotemporal resolution satellite and ground weather data for direct solar irradiance forecasting is used. Despite the Bureau of Meteorology's quality baseline network, there are limitations in continuous long-term validated datasets for DNI and aerosol measurements across Australia. In addition to clouds, the main extinction entity for solar radiation aerosols from bush fires and desert dust storms are frequent events in Australia and can add large additional uncertainties in modelled DNI forecasts. Up until very recently, comparison with detailed surface radiation measurements on a daily basis has not been the focus of NWP development evaluations. Sensitivity analysis of different radiative schemes to high resolution weather data on NOAA's widely used Weather Research & Forecasting model will increase the accuracy of DNI forecast output from NWP models, providing techno-economic benefits to CSP plant operators and industries that require high spatiotemporal shortwave solar irradiance forecasts.

Acknowledegement: This project is related to Australian Renewable Energy Agency's Australian Solar Energy Forecasting System (ASEFS). ASEFS partners include CSIRO, Australian Energy Market Operator, Bureau of Meteorology, University of NSW, University of South Australia & US National Renewable Energy Laboratory.