"Wind Energy, Transmission, and Production Costs: Does Increased Connectivity Help All?", with Harrison Fell
Abstract: We empirically examine the geographic variation in economic impacts of a significant electricity-transmission capacity expansion project, the CREZ project in Texas, on the value of wind generation. Importantly, we extend the scope of canonical two-region trade models in economic theory onto multiple regions. We show that while, as predicted, the two regions directly affected by the expansion, the West and North zones, experience overall falling production costs and gains from trade, the marginal value of wind generation in the Houston zone declines post-CREZ. These effects are such that, across the region as a whole, the marginal value of wind with respect to production costs remains effectively the same before and after CREZ. Additionally, using machine learning classification techniques, we demonstrate that wind generation actually increases the likelihood of relatively higher prices in the Houston zone. These counter-intuitive generation cost and pricing patterns appear to be driven by changing power trade patterns and dynamic production constraints of thermal generators.
"Wind Generation, Inefficiency, and Market Power in Electricity Markets: Counterfactual Prediction with Dynamic Neural Networks for Causal Inference", (in submission)
Abstract: In electricity markets, a forward price premium (persistently positive forward minus spot prices) may arise due to market power and is linked to productive inefficiency. I examine the impact of recent dramatic increases in renewable generation on a systematic forward price premium (FPP) and firms' intertemporal production decisions. Specifically, first, I develop a theory model of imperfect competition to show how increased wind generation can increase the FPP. Next, I implement a novel empirical approach using machine learning methods to predict the impact of wind generation on the FPP and strategic generation behaviors of conventional generators. Consistent with the theoretical prediction, I empirically find that increased wind generation increases the FPP. Also consistent with the theory model, I find dominant firms withhold supply in the forward market to engage in intertemporal price discrimination as wind generation increases. In contrast, fringe producers respond to the FPP increases by increasing supply in the forward market to exploit the price differentials. The results point to the challenges in achieving market efficiency with increased wind generation in the presence of market power. Furthermore, the variety of ways, in which the other factors accompanied by wind generation—forecast uncertainty and wind generation variability—affect market outcomes, highlight the importance of extensive consideration when assessing the costs of renewable generation.
"Reinforcement Learning about Market Dynamics and Optimal Investment in the Electricity Industry"
Abstract: This paper estimates the social benefits and costs of contested policy proceedings aiming to support the efficient investment and performance of procured resources, and further identifies the optimal policy based on the analysis. To do so, I develop a two-stage model to highlight the dynamic nature of electricity suppliers' decision-making. In the first stage, firms make investment decisions with different combinations of generation technologies and storage capacities. In the second stage, based on the capacities installed during the first stage, firms repeatedly open and close operations and adjust their production to maximize profits in the face of fluctuating demand. The second stage incorporates generator startup costs as the entry/exit friction and ramping constraints. Two approaches can further complement the analysis. First, to overcome the computational challenges of solving a complex dynamic competition model, this study employs Reinforcement Learning in Artificial Intelligence to efficiently utilize conventional dynamic programming techniques in economics. The novel method can reduce computational time by more than 70 percent without sacrificing accuracy when solving sequential optimization problems. Second, using this powerful platform, I particularly focus on spatial and temporal heterogeneity in the values of renewable energy and the complementarity of generating technologies by incorporating high resolution wind speed and solar radiation data. Three main results are found from the analysis. First, the presence of dynamic constraints significantly changes the mix of invested technologies and the way to utilize the installed technologies. Second, the current volumetric-based subsidies may excessively incentivize productivity and bias wind investments toward high-producing sites, thereby creating additional social costs. Third, there are significant tradeoffs between consumer and producer surplus for different investment portfolios.
"Optimal Trade-offs between Productivity and Variability in Wind Energy Investments", with Kyle Bradbury and Artem Streltsov
We identify the optimal siting plans of wind generators for each regional electricity market in U.S. by using two modeling frameworks that quantify trade-off between productivity versus variability of system-wide hourly wind generation. First, using high resolution wind speed data at 126,000 sites in the continental U.S., we apply Mean Variance Portfolio theory to compute the share of systemwide wind capacity at each site that yields the highest average hourly power generation for a fixed level of power variability. Second, we train a Long Short-Term Memory (LSTM) network—a machine learning technique suited for time series analysis—to predict counterfactual hourly dispatch curves based on a given level of residual demand, defined as total demand minus production from renewables. The counterfactual predictions using the residual demand data at each region alleviate modeling complexity reflecting grid congestion in reality, possibly more exacerbated with additional wind generators . We next use the predicted dispatch curves to exploit two incompatible changes in electricity production costs driven by each portfolio: (1) the impact of increased wind variability on "out of merit" losses by dispatching higher marginal cost units and (2) a decrease in costs by supplanting more fossil fuel generators with higher levels of wind generation with zero marginal cost. Finally, we compare the cost measures computed by different dispatch curves to identify socially optimal locations of wind generator. Our results demonstrate that the optimal geographic diversification of renewable generators can significantly reduce the variability of renewable generation. More importantly, current wind turbine locations likely driven by volumetric-based government subsidy are biased toward a limited number of high producing sites and create externality.
"Adaptive Resistance Management with Uncertain Fitness Costs", with Zachary S. Brown, (in submission)
We investigate the adaptive management (AM) of pesticide resistance, with uncertainty in parameters governing renewability of susceptibility within the pest population. Our biological model is a multidimensional Markov process determining pest density and frequencies of pesticide-resistant genotypes. By specifying the genotype fitnesses as log-normally distributed, we obtain an analytical posterior distribution of the renewal rate for pesticide susceptibility. Beyond the general contribution of showing how to tractably apply adaptive management to bioeconomic models of pesticide resistance, we present three specific insights. First, we show how to use the multi-dimensionality of the model to optimally weight information from each state variable to hasten learning. Second, in a simulated application to transgenic Bacillus thuringiensis (Bt) corn to control European corn borer, we find that the expected value of perfect information (EVPI) about renewability of susceptibility decreases as the pest population becomes more resistant to Bt. However, the value of information (VOI) from adaptive management (bounded above by the EVPI) increases as the population becomes more resistant. Third, when resistance levels are high, 'active' AM including the VOI within the optimization yields qualitatively greater performance than passively updating beliefs following optimization of the management policy.
"Preventing Dengue using Wolbachia Infected Mosquitoes: Developing Optimal Release Strategies on an Uncertain Time Horizon", with Zachary S. Brown, Brandon Hollingsworth, Alun Lloyd, Julian Sass, and Michael Vella, (in submission)
Dengue virus is a systemic viral infection spread most commonly by the mosquito. Dengue is estimated to infect 390 million individuals a year worldwide, with around 3.9 billion people at risk of infection world-wide. Currently, there are no licensed vaccinations or therapeutic treatments for the disease and control of the vector species through conventional methods has proven inefficient and costly. Recently, there have been several novel vector control techniques proposed to prevent dengue outbreaks. Of these, the approach most likely to be implemented in the near future relies on release of Wolbachia infected mosquito to block the ability of the infected mosquito to acquire and transmit dengue. However, the rearing and release of the infected mosquitoes will require funding and optimal release strategies have yet to be discussed. In this paper, we describe a framework based on dynamic programming by using a Bellman equation to determine an optimal release strategy for the infected mosquitoes in an area at risk for epidemic dengue with an uncertain time horizon. We then use this framework to determine an optimal release surface and find that it is relatively insensitive to most parameters and, in most situations, suggests an "all or nothing" release plan. This allows the optimal release question to be reframed in terms of optimal facility size which is found using fixed cost analysis.
"The Community Explorer: Using County-Level Data on US Diversity Effectively to Inform Policy", with Claude Lopez and Maggie Switek, Milken Institute Report, 2022
"How to Identify Health Innovation Gaps? Insights from Data on Disease Costs, Mortality, and Funding", with Claude Lopez and Brittney Butler, Milken Institute Report, 2021
"An Estimate of Lives Saved by Enforcement of Seat Belt Usage in North Carolina", Institute for Transportation Research and Education Report, 2016
"Risk-based Inspection Resource Allocation Model", Institute for Transportation Research and Education Report, 2016