A multi-scale solar energy forecast platform based on machine-learned adaptive combination of expert systems
Complementary to the above-noted efforts, in this talk, we present a versatile strategy for improved forecasting. Indeed, the prediction of atmospheric state affecting solar energy production, due to its inherent complexity, is unavoidably limited by simplification/coarse graining of the physical system and the uncertainties of boundary conditions. We postulate that the prediction accuracy can nevertheless be improved by exploiting a number of independent prediction methodologies (referred to, hereafter, as expert systems), which involve different sets of simplification/coarse graining. By regressing historical predictions vs. measurements, one can obtain a set of weighting coefficients that by-and-large indicate how well different expert systems work under different circumstances. We further postulate that the weighted combination of expert systems using weighting coefficients derived from historical data may be used to achieve improved forecasting accuracy compared to any individual expert system. Note that the optimal weighting coefficients may be a function of time as well as the meteorological state.
Here we present an implementation of a multi-scale solar forecasting platform which puts the aforementioned hypothesis into live test. The system leverages big-data technology and incorporates atmospheric and cloud predictions from numerous expert systems (sky cameras, satellite imagery products and various NWP models etc.). Machine-learning strategies, similar to those used by IBM Watson in the Jeopardy! Grand Challenge, are used to determine optimal combinations of the different expert systems adaptively depending on the categorization of meteorological states and other factors. Preliminary results empirically showing the advantage of such strategy will be presented. The experience of applying such adaptively-combined-expert-systems in solar forecasting indicates the strategy may have far-reaching potential for the simulation and prediction of other complex physical systems.
Acknowledgement: The work is partially funded by the DOE SunShot Initiative.