At the inception of electricity markets at the turn of the millennium, stochasticity in power systems was more readily manageable. Electric demand followed the rhythm of daily human activity, and load forecasting systems could forecast demand quite accurately by taking into account weather conditions. Unit outages were random events which were not correlated with each other. As were transmission outages. The generating units were fossil fuel fired with abundant fuel supply and could be turned on (committed) and dispatched to follow electricity demand with relative ease. The limited stochasticity in the grid was managed by carrying enough operating reserves to meet sudden loss of units (typically by carrying a multiple of the largest unit in the system), and the grid was operated in a “secure mode” to withstand the loss of one major transmission element (N-1 condition).
This manageable state of affairs has been upended in the last 25 years. The grid conditions are far less predictable. Rapid penetration of large quantities of intermittent renewable generation requires grid operators to ramp down fossil units when the sun is shining, or wind is blowing and manage sudden fluctuations of renewable output by utilizing storage resources or standby generation (reserves). The fact that wind or solar generation can be highly correlated in a geographic region adds to the problem. For example, large swaths of wind plants can suddenly drop in output. Demand used to follow a predictable pattern, but proliferation of micro-grids, distributed resources, demand response, “behind the meter” aggregators and storage resources now means that demand can be highly volatile and less predictable. Decarbonization is also leading to rapid electrification of entire economic sections such as transportation (electric cars) and heating (heat pumps) and adding to the stress on the power system which is struggling to meet demand and manage uncertainty. The Artificial Intelligence (“AI”) revolution involves consumption of prodigious amounts of power by server farms that are at the heart of the AI learning models. AI is adding to the demands on a stretched grid powered increasingly by intermittent and distributed resources. All this is making life increasingly difficult for grid operators and planners tasked with the responsibility to “keep the lights on” through all this uncertainty.
Sadly the industry tools have not evolved fast enough to keep up with the increased stochasticity in electric grids. But progress is being made. For example, several wholesale market operators are implementing advanced AI and machine learning based tools to forecast renewable generation and electricity demand. Several grid operators are also developing the capability to set operating reserves dynamically based on system conditions such as the level of renewable generation and transmission headroom available. However, the core unit commitment algorithms which commit and dispatch available resources as well as the industry standard system planning tools still use deterministic assumptions or an incomplete representation of system stochasticity. Much more needs to be done to model the stochastic nature of today’s grid.
Loudon Energy Analytics is hoping to be a catalyst for introducing comprehensive stochastic modeling for grid operators, planners and traders. We are commercializing the ORFEUS Software platform developed by researchers at Princeton University’s Operations Research and Financial Engineering (ORFE) Department which received over $4 million in funding as part of DOE ARPA-E’s Perform initiative. This research uses a full fundamental representation of the grid, incorporates stochastic scenario development and uses financial engineering concepts for stochastic optimization. This system can optimize based on risk, environmental profile or market revenues and find the right trade-off between competing objectives. We are eager to work with grid operators, planners and power marketers to apply this tool to real world problems.