In this research, we focus on the strategic interac-tion among a number of power generation compa-nies (generators) at a generic trading floor in an electricity market (which can be exemplified with the day-ahead market). Such a trading floor is ob-served in a deregulated the power industry; each generator submits bids to generate power periodi-cally, e.g. for every hour of the following day in a day-ahead market. The Independent System Operator (ISO) collects the bids, and solves an optimization problem to determine the power to be dispatched from each generator, and the electricity prices at different locations in the power grid.
We aim to understand how the collective bidding behavior of generators evolves over time and if it particularly evolves to an undesirable outcome such as a collaborative collusion among the generators. This is an important subject of inquiry because the power dispatched from, and the profit of every generator depends on the bids of all generators in the market. Based on existing work in literature (such as Krause et al. 2004, Weidlich 2008, Bakirtzis and Tellidou 2006 ) we have developed an agent-based simulation model where the generators are modeled as adaptive software agents. The agent-based simulation approach offers the flexibility required to model this dynamic market, for instance, the ability to model heterogeneous generators with different production capacities, marginal costs, and behavioral parameters. Such flexibility is often not possible with traditional analytical models.
The agent-based approach also allows us to model the learning behavior of generators. The price to bid is not a trivial choice for a generator, because the best bid for a particular period of a repetitive trading floor depends on the bids of other generators in the system. The generator knows neither its competitors' bids nor their marginal costs or production capacities or financial situation. However, since the bidding floor is repeated continuously, generators may learn through experience the effect of different bids on their earnings. To this end, we use a reinforcement learning approach and model generators as adaptive agents who learn from past experience and adapt to environment.
Analyzing simulation data, we characterize how the generators' bids, learning model dynamics and re-sulting electricity prices evolve over time. We study various questions under different market, network and learning scenarios. Some generators may have advantage over others due to lower production costs or due to favorable location: Will these gener-ators discover their advantage and benefit from it? A group of generators may increase their profits when they all place high bids: Will such groups of tacit collusion emerge among generators? Game-theoretical models of the system predict certain Nash Equilibrium in bids: Does the simulation with active agents converge to such equilibria? The Inde-pendent System Operator may choose between various market rules, such as pay-as-bid pricing versus uniform pricing. Which pricing rule leads to a more competitive market? How would the physical characteristics of the power grid, such as the trans-mission line constraints, or the behavioral characteristics of generators, such as their risk aversion levels, affect the results?
We plan to extend our initial model to study more realistic, hence more complicated power market settings. One extension is to consider other markets beyond the day-ahead market, such as the balancing market. Yet another extension is to include demand-side agents in the model. To approximate a real power network, we may also increase the size of the considered market.
Ultimately, our findings can help governments de-sign more efficient power markets, leading to lower cost electricity generation and increase in public welfare. Allowing what-if scenario analyses, such simulation models can be used as a "wind tunnels" to test the effects of policy proposals on market players. They can also be employed to address greenhouse gas emissions of generators.