Key takeaway: A median weather year is selected after analysing decades of weather data, balancing solar, wind, temperature, and outage patterns across the NEM.
Certain components of the model rely on a designated “weather year” - a representative historical period selected as a seed to inform projections of future weather conditions and associated outage profiles. This is used to seed our demand, solar/wind load factors, hydro, generation outages, and transmission outages.
Using one weather year makes the model sensitive to the weather year chosen. To make sure a representative weather year was selected, a sensitivity analysis of the last 10 years was conducted. Years were evaluated based on solar and wind traces, temperature, number of generation and transmission outages, and geographical distribution of those outages. 2023 was chosen as it was the closest year to the median across all metrics.
Alternative approaches considered
For full transparency, we also considered and discounted the following options for seeding the above:
Cycling between 10 weather years: this is a method that AEMO use in some of their modelling, but we discounted it as it then becomes unclear whether a year in the forecast has high revenues due to fundamentals or due to using a weather year with a high number of outages.
Randomness based on historical distributions: this sounds like a clever way of modelling things, but we believe strongly against adding any deliberate randomness to the model. It only makes the model more of a black box, and it’s hard/impossible to capture this randomness perfectly (especially correlation between demand/solar/wind). Also, unless we repeat the randomness every year, this approach still has the downside of making it unclear whether a particular year has high revenues due to fundamentals or due to randomness.