We define “medium-range” to be the horizon beyond short-range and out as far as you might reasonably expect underlying trends, like population and the economy, not to shift significantly, generally about one year. Anything beyond that we consider “long-range.” Either way, we take the same approach to projecting demand.
All TESLA Models estimate a latent trend variable that is designed to capture the slow growth or decline in the base load of a system from month to month and year to year. For medium- and long-range projections, we first project this trend variable across the desired horizon allowing us to inflate or deflate our projections as we expect base load to grow or decline. This process makes no explicit assumptions about economic or population growth nor gains in energy efficiency or demand response programs. We believe this represents a “business-as-usual” assumption. While this is our default assumption, we can manipulate the trend variable to force more explicit assumptions.
Monte Carlo Weather Simulation
The TESLA Model is dependent on a weather forecast input. For medium- and long-range horizons, we employ a Monte Carlo simulation using long histories of actual weather data to generate that weather input. This process involves forecasting future periods using historically observed weather data. By shifting historical data a day at a time to force different weather patterns onto different days of the week, we can generate multiple different demand projections from a single year of historical weather data. When repeat the process with the next year of historical weather data.
Running the TESLA Model against a wide array of weather inputs results in a wide array of demand projections, generally we will generate 150 projections given a ten year weather history. From that array, we are able to extract and present the 10th, 20th, 50th, 80th, and 90th percentiles along with a maximum projection.