Probabilistic Net-Load Forecasting Using Machine Learning

Presentation Description: Probabilistic forecasting of net load (electrical demand less renewable generation) has been slow to be fully adopted by utilities and grid operators. Skill metrics for probabilistic forecasts are more complicated than for deterministic forecasts, and forecasts of the tails of the probability distribution require much longer training timeseries to confirm that rare events happen with the predicted frequency. These long training samples typically require synthesizing generation and load data, since generation and demand infrastructure change over time, but this synthesis itself introduces hard-to-quantify errors to the forecast process. Finally, load forecasting and renewable generation forecasting have typically been performed by different vendors, complicating the production of a probabilistic forecast of their difference.
To overcome these difficulties and demonstrate the potential costs savings of skillful net-load forecasting, we are participating in research funded by the US DOE ARPA-E program to develop decision support in the form of risk tranches for renewable generation. This presentation will review our experimental set-up for simulated electrical grids in New York and Texas, which will allow us to demonstrate the usefulness and reliability of the probabilistic forecasts of net load in systems with a range of renewable generation penetration.