Planning a reliable electricity system means ensuring supply can efficiently meet the demand of growing populations and economies, as new loads come online, and investment flows into a widening range of generation technologies. Many options exist for meeting that demand and choosing among them requires understanding how different sources of generation will interact in future scenarios. That calls for models that not only reproduce the past accurately but stay flexible enough to accommodate new technologies, market mechanisms, and strategies as they emerge.
Supply-side modelling divides into dispatchable generators, whose output is controlled directly, and variable generators, whose output is primarily dictated by weather. Wind turbines and solar panels follow the rules of physics; feed weather data through a power curve or an irradiance model, and the output matches historical time patterns and distributions, after accounting for some noise and bias [1, 2]. We fit the physical relationships to historical data, then let weather and climate projections drive the resulting generation.
Dispatchable generation is controllable, but how it’s controlled introduces an array of new variables. Maintenance, profit strategies, portfolio size, and operator preferences all shape what reaches the market [3]. Physics still plays a role, though: converting fuel to power follows thermodynamic laws, and we can estimate each unit’s costs and limits. The challenge is deciding which relationships can be described directly from physics, and which are better learned from how generators have actually operated.
Single driver, single source of truth
Take a hot afternoon with sunny skies. Gas turbines breathe air, and hot air lowers their maximum capability [4]. This is called a derate. The same heat raises cooling demand, and the sunny skies push solar output toward its peak. Solar fades by early evening, while the heat, the demand, and the derate linger. Weather moves conventional supply, demand, and renewables together, so a simulator should use the same weather conditions to drive all three. In system-wide data, though, thermal derates can hide among units on outage or already running at part-load. A model fitted only to generation data may miss the link between a derate and a hot day.
Physics tells us the derate arrives with the heat, often when the system is at its tightest. By building known physical relationships into the simulator, we preserve those connections while using data-driven models where the underlying behaviour is harder to specify. The result is a model that is both interpretable and flexible: it shows what a generator does, which factors drive it, and what happens when we alter those factors [5].
The merit order and market dynamics
Modelling the physical behaviour of each generator is only part of the problem. Electricity systems are also markets, and market behaviour determines which generators actually operate. Outcomes depend on the merit order, where all offers of price and capacity are sorted and dispatched to meet changing demand. Under Alberta’s current rules, each generator can submit up to seven offer blocks, each with a price and a quantity [6]. With new market rules on the horizon under Alberta’s Restructured Energy Market (REM) [7], offers will interact in new ways, producing scenarios and outcomes absent from the historical record.
Understanding every individual offer is impractical, but capturing how offer strategies change is valuable. We therefore represent each generator’s offer using a small number of explanatory parameters: availability, capacity offered at $0, capacity held back at the price cap, and a fitted curve between them. Like the physics-based components, each parameter has a clear interpretation. The approach can reproduce historical price and dispatch patterns while also giving the user a dial to explore alternative strategies. For example, a user could toggle a generator from “conservative” to “aggressive” and watch how prices respond, or add a new unit and see how the market reacts.
The simulator should first demonstrate stability and accuracy against historical data. But its real value is in exploring scenarios beyond the historical record: testing the capability of the current fleet against new loads, bringing new generation online, developing previously untapped renewable resources, or running the fleet under REM rules. Comparing options side by side, with a clear view of why each performs as it does, turns a simulation into better-informed decisions.
Interested in discussing further? Connect with Matt on LinkedIn.
References
[1] J. Olauson, “ERA5: The new champion of wind power modelling?”, Renewable Energy, vol. 126, pp. 322–331, 2018.
[2] S. Pfenninger and I. Staffell, “Long-term patterns of European PV output using 30 years of validated hourly reanalysis and satellite data”, Energy, vol. 114, pp. 1251–1265, 2016.
[3] D. P. Brown and D. E. H. Olmstead, “Measuring market power and the efficiency of Alberta’s restructured electricity market: An energy-only market design”, Canadian Journal of Economics, vol. 50, no. 3, pp. 838–870, 2017.
[4] F. J. Brooks, “GE Gas Turbine Performance Characteristics”, GE Power Systems, GER-3567H, 2000.
[5] C. Rudin, “Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead”, Nature Machine Intelligence, vol. 1, pp. 206–215, 2019.
[6] AESO, ISO Rules Section 203.1, Offers and Bids for Energy.
[7] AESO, “Restructured Energy Market Final Design”, August 2025.
