Agentbased Models for Electricity Policy Design Derek W

Agent-based Models for Electricity Policy Design Derek W. Bunn Augusto Rupérez Micola London Business School

Background o Series of 5 Ph. D dissertations at LBS since 1994 applying ABS to competitive pricing behaviour in electricity markets o Each addressed Specific Questions for sponsoring companies and government inquiries – Normative company insight for business strategy and risk management – Normative institutional insight into market reforms and regulation o What follows is a selective, critical review 9/17/2021 2

Business Strategy Questions o How do players with substantial market power maximise returns over time in both spot and forward contract markets? o How can a company understand the new trading risks for an impending spot-market rule change? o As market structures evolve through M &A activities: How should companies position themselves horizontally and vertically? 9/17/2021 3

Institutional Design Questions o How much divestiture should be required of dominant incumbents? o Will a market rule change from a day-ahead Pool auction to Continuous Bilateral Trading impede the exercise of market power? o Does a particular generator have the ability to move the market? o What is the effect of an increasing tolerance for crossholdings and vertical integration? 9/17/2021 4

Why Should ABS Be More Useful? o Electricity markets are too complex and the required answers too subtle for analytical methods: – – – Imperfect competition Very low demand elasticity to price Discontinuously convex supply functions High-frequency repeated game Heterogeneous agents • • 9/17/2021 Size Portfolios Technologies Cross-Ownerships 5

Typical ABS Models of the Electricity Wholesale Market Start with detailed representation of the generators’ marginal-cost supply function 9/17/2021 6

Supply Functions and Dynamic Learning: Offers are not at Marginal Cost 9/17/2021 7

Typical ABS Models of the Electricity Wholesale Market o Artificial Agents similarly learn to make offers to the market above their marginal costs o The demand side is represented via a function or more explicitly by a double-sided auction o We have found that both best-response and simple reinforcement learning work well, and can produce credible results 9/17/2021 8

Partial Best-Response Learning and Supply-Function Evolution 9/17/2021 9

Validation Via Theoretical Supply. Function Equilibrium 9/17/2021 10

Model Calibration: Steep Demand Function o Whether ABS or analytical, avoiding very high prices in profit-maximising models of electricity auctions is a pervasive problem o In reality high price spikes occur at times of scarcity, but most of the time generators do not exert their full market power to raise prices – A “long-term” elasticity is often used, as a calibrated surrogate for the repeated game 9/17/2021 11

Local Search Works Better Global Search => Cycles 9/17/2021 12

Divestiture and Contract Cover 9/17/2021 13

Calibration through Utilisation In order to use Realistic Short-term Elasticity —another “calibration parameter” is needed A “Utilisation” factor can be introduced as a target alongside daily profit maximisation This has been framed as a corporate, market-share objective in several reinforcement-learning models (market-power studies for UK, Germany, Italy, Russia) And as regulatory, availability constraint in a best-response learning model for the UK. 9/17/2021 14

Profits and Utilisation 1. Market Share Objective YES Did EACH plant reach target utilisation rate at time T -1 ? 2. Profit Objective YES REPEAT All T -1 bids by random % 9/17/2021 TOTAL profit at time T - 1 exceed that at time T -2 ? NO NO LOWER Plant’s T 1 bid by random % CHANGE All T -1 bids by random % 15

One approach is to see if historical data on utilisation allows the model to fit…. . Mean Simulated and Actual SMP (1 December 1997) 9/17/2021 We then used this model for UK market rule changes 16

Prices were Higher under the Bilateral Model than in the Pool versus Bilateral Model Simulated-Clearing Prices 9/17/2021 17

Another approach is to calibrate the model to historical data via utilisation…. . 9/17/2021 We then used this model for German mergers 18

The Effect of Four Big Mergers 9/17/2021 19

An increase in Target Utilisation Rate increases Agent Rivalry, and Prices Fall 9/17/2021 Which is why incumbents are reluctant to sell plant to new entrants 20

Reserve As Utilisation An equivalent effect follows from reserve capacity We used this in an antitrust-case model Whether two “minor” players could move the market price by capacity withdrawals Using a double-sided reinforcement model Evidence Individually they could not Collusively they could 9/17/2021 21

ABS in Adversity • As a basis for expert testimony, ABS results can be fragile: – Statistical and equilibrium uncertainties can be easily challenged (lawyers don’t like t-tests) 45 GW demand, full capacity 9/17/2021 22

Market Structure and Ownership o Power plants and energy companies are actively traded – This “market” determines the behaviour in the spot o The effects of horizontal concentration amongst generators have been wellresearched o Recently we have used ABS to look at: - Cross-holdings - Vertical relationships 9/17/2021 23

Crossholdings and Coordination P 9/17/2021 Cross-holdings 24

Collusive Value of Transparency PPUB – PPRIV 9/17/2021 Cross-holdings 25

Gas-Power Value Chain Gas Shipping Electricity Generation Electricity Retailing Modelling Two Simultaneous Spot Markets Presents an extra modelling challenge: Sequencing the Clearing 9/17/2021 26

A Trading Perspective on Clearing SG SE PE PG DG QG = Q E = Q R Wholesale Gas Market 9/17/2021 SR PR DR DE QE = Q R Wholesale Electricity Market QR Retail Electricity Market 27

Making Agent-based Models Work: Some Open Questions • Parameterisation for Energy Policy – Role of elasticity • Long-term / short-term – Excess capacity and capacity utilisation – Auction type • Double- / single-sided – Market clearing in sequential markets • Supply chain / Trading netback / Simultaneous 9/17/2021 28

References o Competition Commission. 2001. AES and British Energy: A report made under section 12 of the Electricity Act 1989. http: //www. competitioncommission. gov. uk/reports/2000. htm#2001 o Divestiture of Generation Assets in the England Wales Electricity Market: A Computational Approach to Modelling Market Power. (with C. Day) Journal of Regulatory Economics, 2001 o Imperfect competition in uniform-price and discriminatory auctions for electricity. (with J. Bower) Journal of Economic Dynamics and Control, 2001 o Agent-based Simulation: Modelling the New Electricity Trading Arrangements of England Wales. (with F. Oliveira) IEEE Trans on Evolutionary Computation, 2001 o A Model Based Comparison of Strategic Consolidation in the German Electricity Industry. (with J. Bower), Energy Policy , 29 , pp 987 -1005, 2001 o Current working papers at www. london. edu/ds 9/17/2021 29