Learning spatiotemporal dynamics in wholesale energy markets with dynamic mode decomposition

Abstract

Energy markets balance electricity generation (supply) and demand while ensuring non-discriminatory access. Understanding energy market dynamics is essential to improving grid efficiency and optimizing the development of new energy conversion and storage technologies. Accurate energy price forecasts are important for many energy storage technologies to be profitable from price arbitrage. In this paper, we apply Dynamic Mode Decomposition (DMD), a popular spatial-temporal reduced-form modeling technique, to forecast 6587 locational marginal prices in the California Day-Ahead Market (DAM). DMD is a promising equation-free modeling technique in systems with inherent periodic tendencies in time such as financial markets and fluid dynamics. Yet we show, for the first time, that DMD cannot reliably forecast day-ahead energy prices due to the periodicity of the data. Instead, we show Augmented DMD (ADMD) overcomes these limitations and is a fast and accurate price forecaster. We benchmark DMD, ADMD, and backcasting forecasting methods for optimal price arbitrage with an energy storage system. Using ADMD, a market-connected energy storage system can capture up to 92% of allowable revenues in rolling horizon simulations. Moreover, ADMD is up to 1000-times faster than time-series forecasting methods and requires orders of magnitude less data than deep/machine learning techniques.