Reaching New Lows? The Pandemic's Consequences for Electricity Markets.

• Author: Benatia, David

1. INTRODUCTION

   Under the "Great Lockdown", demand for energy has plunged in the worst economic downturn since the Great Depression (IMF, 2020). Many electricity systems have experienced unusual patterns of consumption, unprecedented demand reductions and coincidentally low wholesale prices due to low commodity prices and drastic public health measures like stay-at-home orders, curfews, and travel restrictions. Electricity demand reductions during the first lockdown episode, in the spring of 2020, have been estimated from 3-4% in Texas (ERCOT, 2020) up to 20-25% in Italy (Narajewski and Ziel, 2020). Those large sudden variations are unparalleled in history, even during major economic crises (RTE, 2020). (1)

   The first case of COVID-19 in New York, the virus epicenter in the United States in the spring of 2020, was reported on March 1, 2020. Governor Andrew Cuomo declared a state of emergency on March 7, and schools closed in New York City (NYC) on March 16. The statewide stay-at-home order became effective on March 22 in an effort to contain the outbreak, with above 4500 daily new cases. Public health policies have changed many times in the following months in attempts to reopen the economy while managing sanitary risks. This paper finds that the pandemic has set new record lows in New York's electricity markets and undermined system efficiency. Similar measures are likely to have had comparable consequences in many other electricity markets across the globe.

   This paper has three main contributions. First, it provides a detailed empirical analysis of the pandemic's consequences on New York's electricity markets. Second, it develops a methodology to evaluate the impacts on electricity demand using machine learning, namely a two-layer feed-forward neural network. The key advantage of this method lies in its ability to yield precise hourly counterfactual predictions using only exogenous factors such as weather and time variables. Third, it builds an empirical framework to disentangle the respective short-term effects of demand reductions, increased forecast errors, fuel price drops, and the coincidental retirement of a nuclear reactor on wholesale prices in the day-ahead and real-time markets. A standard reduced-form approach evaluating the pandemic's consequences would fall short in disentangling these effects. Instead, this framework consists in estimating a structural econometric model of the market-level supply, for both day-ahead and real-time, and then combining it with counterfactual predictions obtained from the neural network and regression models. This method allows simulating counterfactual hourly price scenarios under various assumptions about market fundamentals such as demand, forecast errors, fuel prices, and nuclear capacity. The estimates are finally used to assess changes in revenues for market participants and additional operating costs caused by unusually large forecast errors. We conclude the paper by discussing some implications for market participants, grid operators and end-users.

   The empirical framework based on machine learning counterfactual predictions used in this paper was inspired by the work of Burlig et al. (2020) on energy efficiency. There is a burgeoning literature in energy and environmental economics using machine learning methods for policy evaluation and regulation (Abrell et al., 2019; Benatia and Billette de Villemeur, 2019; Benatia, 2022; Fabra et al., 2022; Graf et al., 2020). This paper combines a machine learning approach with a structural econometric model to obtain counterfactual market outcomes assuming the pandemic had not occurred.

   There are also many papers studying the impacts of the COVID-19 pandemic on energy markets. Ftiti and Louhichi (2020) review the dynamics of the oil markets during the pandemic and discuss the prospects for the industry moving forward. Tsagkari (2020) argues that the pandemic will radically change the energy sector as demand shrinks and the digitization of the economy develops further. Cicala (2020) makes use of electricity demand reductions to proxy the slowdown of economic activity. (2) Focusing on the energy transition, Gillingham et al. (2020) argue that the delays in renewable capacity additions and innovation caused by the crisis might outweigh the emissions reductions of the first lockdown.

   The effects of sanitary measures on electricity demand and forecasting performance have also attracted much attention from power system operators and academics (AESO, 2020; Benatia, 2022; Brewer, 2020; ERCOT, 2020; Graf et al., 2020; Narajewski and Ziel, 2020; Percy and Mountain, 2020; PJM, 2020; NYISO, 2020; RTE, 2020). (3) Benatia (2022) focuses on the lost revenues for suppliers, retailers and system operators in France and discusses the lessons learned for the energy transition. Percy and Mountain (2020) finds that demand reductions are consistent with reductions in mobility following the imposition of social distancing regulations in many regions. The high-dimensional change-point model of Narajewski and Ziel (2020) reveals significant changes in daily and weekly consumption patterns in continental Europe, especially in France and Spain. Brewer (2020) shows that PJM's day-ahead load forecast error jumps to 3% on average which corresponds to the worst performance in recent history. Graf et al. (2020)'s deep-learning model reveals that real-time balancing costs have doubled on average in Italy during the first lockdown, and use their estimates to evaluate the balancing needs throughout the energy transition. To the best of our knowledge, this paper is the first to: 1.) study the effects of the crisis during the entire first year; (4) 2.) account for spatial heterogeneity in demand reductions; and 3.) identify the separate effects of market fundamentals on balancing costs as well as day-ahead and real-time markets.

   Our analysis highlights three main empirical facts. First, it reveals heterogeneous electricity demand reductions across regions in New York during the year. Statewide reductions are estimated at 4.6TWh (-3%) from March 2020 to February 2021, with respect to the counterfactual demand assuming the pandemic had not occurred. New York City (NYC), representing a third of total electricity consumption, was the most affected with reductions corresponding to 85% of the statewide net effect. In opposition, electricity consumption in some other demand zones have been persistently larger than usual after the first lockdown episode. (5) Daytime consumption has decreased by 17% in NYC and by 10% statewide during weekdays and daily consumption patterns have considerably changed during the first 6 weeks of lockdown. This spatial heterogeneity is most likely explained by: a.) large outflows of urban residents to less-populated areas, (6) and b.) differences and changes in shares of residential, commercial and industrial consumption across zones. Those estimates are in line with EIA (2020) and NYISO (2020), and are qualitatively similar to other electricity systems in North America (AESO, 2020; Brewer, 2020; ERCOT, 2020; ISONE, 2020; PJM, 2020). Demand reductions in Europe have, in comparison, been significantly larger due to more stringent sanitary measures (Benatia, 2022; Fabra et al., 2022; Narajewski and Ziel, 2020; RTE, 2020).

   Second, average hourly balancing needs have increased by 72% from March to July, 2020 due to unusually large short-term load forecast errors, leading to an over-reliance on the real-time market. The algorithm for short-term load forecasting in New York failed to quickly adjust and has resulted in both chronic over-forecasting and more volatile errors for months. Around the globe, system operators have mobilized their workforce to attenuate forecast errors in an effort to mitigate its economic consequences (NYISO, 2020; RTE, 2020). Large forecast errors result in inefficient daily system operations because of additional operating costs from unnecessary start-ups and provisions of spinning reserves (Ortega-Vazquez and Kirschen, 2006). On one hand, over-prediction wastes resources since more reserves are available than actually needed, hence increasing operating costs. On the other hand, under-prediction increases operational costs because the additional energy must be procured in real-time from more expensive units (Hahn et al., 2009; Kyriakides and Polycarpou, 2007). This paper finds that balancing costs in New York have increased by $12 million (+46%) in the first five months of the crisis, due unusually large load forecast errors, resulting in significant losses of system efficiency. These estimates neglect the associated environmental externalities.

   Third, wholesale prices have dropped by $3.3/MWh (-15%) in the day-ahead (DA) market and $5/MWh (-23%) in the real-time (RT) market, on average during the first months (March-July 2020). (7) The variations in day-ahead (DA) prices are attributed to demand reductions (43%), low fuel prices (25%) and the retirement of Indian Point Unit 2, a nuclear power generator of 1032MW, over the same period (32%). DA prices would have been smaller by on average $1.5/MWh had the reactor not retired. (8) On the other hand, real-time (RT) prices were also affected by increased forecast errors (17%), in addition to low demand (36%), low fuel prices (21%) and the reactor's retirement (26%). Combining the above estimates implies a depreciation of the DA market of $250 million (-6%) but an appreciation of the RT market of $15 million (+23%) over the entire year.

   Therefore, the pandemic's consequences for New York's electricity markets have been large, heterogeneous, driven by multiple factors, and quite persistent during the first year. For end-users, reductions in demand and electricity procurement costs did not translate in savings on energy bills. Residential consumption has increased by 6.7% in New York, most likely due to telework (Cicala, 2020), and the average...