Can agent-based models forecast spot prices in electricity markets? Evidence from the New Zealand electricity market

Modelling price formation in electricity markets is a notoriously difficult process, due to physical constraints on electricity generation and transmission, and the potential for market power. This difficulty has inspired the recent development of bottom-up agent-based algorithmic learning models of electricity markets. While these have proven quite successful in small models, few authors have attempted any validation of their model against real-world data in a more realistic model. In this paper we develop the SWEM model, where we take one of the most promising algorithms from the literature, a modified version of the Roth and Erev algorithm, and apply it to a 19-node simplification of the New Zealand electricity market. Once key variables such as water storage are accounted for, we show that our model can closely mimic short-run (weekly) electricity prices at these 19 nodes, given fundamental inputs such as fuel costs, network data, and demand. We show that agents in SWEM are able to manipulate market power when a line outage makes them an effective monopolist in the market. SWEM has already been applied to a wide variety of policy applications in the New Zealand market.2