UC Riverside

The wholesale electricity markets in the United States operate under the two-settlement system, comprising the day-ahead (DA) market and the real-time (RT) market. The DA market clears bid-in supply against bid-in demand and determines DA physical schedules for generators, virtual awards, and DA locational marginal prices (LMPs), which are defined as the marginal costs of serving the next increment of demand at pricing nodes consistent with existing transmission constraints and performance characteristics of generation resources. The RT market procures "balancing" energy to meet the forecast RT grid energy demand and determines RT dispatch signals for resources and RT LMPs. The introduction of virtual bids to electricity markets is to mitigate the discrepancy between the DA market and the RT market in LMP spreads. In this dissertation, we focus on developing a machine learning-driven portfolio optimization framework for virtual bidding in electricity markets considering both risk constraint and price sensitivity. The algorithmic trading strategy is developed from the perspective of a proprietary trading firm to maximize profit. A recurrent neural network-based LMP spread forecast model is developed by leveraging the inter-hour dependencies of the market clearing algorithm. The LMP spread sensitivity with respect to net virtual bids is modeled as a monotonic function with the proposed constrained gradient boosting tree. We perform a comprehensive empirical analysis on PJM, ISO-NE, and CAISO with the proposed framework. We further introduce a similar framework to arbitrage congestion with virtual bids. A deep neural network is designed to estimate the difference between congestion spreads in DA and RT markets. A clustering algorithm is adopted to separate pricing nodes into a few groups, between which the congestion spreads can be exploited. We validate the proposed algorithmic trading strategy using publicly available data from CAISO.

The electric power system is a major contributor to greenhouse gas (GHG) emissions. To reduce GHG emissions, accurate emission predictions are essential. The marginal emission factor (MEF) is a useful signal for distributed energy resource aggregators and end-use customers to mitigate GHG emissions by scheduling the flexible loads accordingly. We propose a hybrid machine learning framework to predict GHG emissions and locational MEF, which integrates feed-forward neural networks (FNNs) with spatio-temporal graph convolutional networks (STGCNs). A comprehensive case study in CAISO shows that the proposed approach outperforms the existing techniques in prediction accuracy. The proposed model provides short-term locational MEF predictions with high time granularity using only publicly available data.