Financial Arbitrage and Efficient Dispatch in Wholesale Electricity Markets

Transcription

1 Financial Arbitrage and Efficient Dispatch in Wholesale Electricity Markets John Parsons, Cathleen Colbert, Jeremy Larrieu, Taylor Martin and Erin Mastrangelo MIT Sloan School of Management & MIT Center for Energy and Environmental Policy Research FERC and CAISO Harvard Energy Policy Seminar, September 28, 2015

2 Motivation Specific: Problems with Virtual Bidding General: Controversy Surrounding Financial Trading in Commodities Does speculation disrupt the market in the physical good? Or, does it improve the market s functioning? A Unique Opportunity to Analyze an Engineered Financial Market

3 Convergence of Zonal Energy Prices DA/RT Spreads ble 3 evaluates price convergence at the zonal level by reporting the percentage difference ween the average day-ahead price and the average real-time price in select zones, as well as average absolute DA/RT value of spread the difference = Day-Ahead between LMP hourly day-ahead Real-Timeand LMP real-time prices from 11 to Table 3: Price Convergence between Day-Ahead and Real-Time Markets Table 1. DA/RT SpreadsSelect in Select Zones, Zones of the NYISO, Avg. Diff % (DA - RT) Avg. Absolute Diff % Zone West 1.4% 0.0% -1.9% 24.0% 26.4% 36.3% Central 1.1% 0.6% 1.3% 25.7% 25.5% 29.5% Capital 2.6% 2.9% 4.5% 28.1% 27.0% 33.1% Hudson Valley 0.9% 0.9% -0.8% 30.0% 29.9% 33.9% New York City 1.8% 0.8% -1.4% 32.4% 31.4% 35.0% Long Island 0.9% 1.7% -6.5% 35.5% 42.1% 46.5% measured by the average difference in day-ahead and real-time prices, energy price vergence was fair in most areas in Convergence in 2013 was worse than in 2012.

4 Virtual Bidding Day-Ahead market. Like any other bid, a (p,q) pair for any hour and for some node/zone, but flagged as virtual. Bids clear like all other bids. Gross payoff to cleared bid equal to the DA/RT spread. π VD,i,t = (RT i,t DA i,t ) (1) Other charges can materially affect the net payoff. Directly impacts unit commitment and dispatch schedule in the Day-Ahead market. Intended to produce convergence, i.e., reduce DA/RT spread.

5 Apparent Success CAISO Hourly DA/RT Spreads, Before and After Virtual Bidding AFTER DA/RT Spread, $/MWh BEFORE Hour of the Day

6 The Fault with Virtual Bidding

7 Received Theory: Complete Markets, Run Successively Each market is a complete solution. Successive runs of the same market design. The only difference is information. Thought experiment: Given the same supply and demand, they will produce the same dispatch and prices. Day-Ahead is a often described as a cash settled forward market. Equilibrium forward price must be the expected spot price-with a minor risk adjustment.

8 Reality: Approximate Solutions, Articulated The combined unit commitment and optimal power flow problems are too complex to solve as needed. Sensible unit commitment requires looking far forward in time, which requires many hours of computing time. Transmission constraints are highly non-linear. Dynamic stochastic optimization is far too large a problem. Even 50 years after the problem was first formulated, we still lack a fast and robust solution technique for the full alternating current optimal power flow problem. We use approximations, decompositions and engineering judgment to obtain reasonably acceptable solutions to this problem.

9 Reality: Approximate Solutions, Articulated (2) Multi-settlement market design is a strategic solution break the problem into two articulated steps. Day-Ahead market does unit commitment and a granular dispatch schedule, but with a simplified representation of the transmission system and various other constraints. Real-Time market is better resolved, but takes the unit commitment and granular dispatch schedule as the starting point. A miscellany of other constraints are imposed to satisfice on the optimal stochastic solution. Day-Ahead is not just a cash settled forward market. Results of the Day-Ahead market are passed to the Real-Time market and shape that outcome.

10 Reality: Approximate Solutions, Articulated (3) Works reasonably well. Until it doesn t. Over time, system parameters change investments in new generation, new technologies, new transmission and context changes, too prices of fuel, emissions penalties and constraints, etc. Recalibration. Requiring adaptation in the approximations, decompositions, and engineering judgments employed in order to approximate the optimal solution. Convergence should be understood as a practical diagnostic. DA/RT spreads are a product of the approximations and articulated design.

11 Example: Black Start & Voltage Support in PJM PJM West PJM East Load Load CT CT Coal CT Coal Coal Coal CT CT CT CT Load Coal Load Load

12 Case Study of Virtual Bidding in California

13 Average Prices 2009, Pre-Virtuals Two problems: 1 Price spikes in RT, causing negative average DA/RT spreads. 2 Low Hour-Ahead prices, causing positive average DA/HA spreads.

14 Offsetting Virtual Bidding Strategy RT Spikes mean virtual demand at internal nodes is profitable. π VD,i,t = (RT i,t DA i,t ) (2) Low HA price means virtual supply at interties is profitable. π VS,j,t = (DA j,t HA j,t ) (3) An offsetting pair strategy is even better: π VD,i,t + π VS,j,t = (RTi,t HA j,t ) (4) Profits from both pricing problems. Less risk eliminates the uncertainty in the accuracy of the Day-Ahead price. Net zero change in the Day-Ahead demand/supply. That is, zero social welfare improvement from dispatch. But, a parasitic drain to virtual bidders.

15 2/1/2011 2/15/2011 3/1/2011 3/15/2011 3/29/2011 4/12/2011 4/26/2011 5/10/2011 5/24/2011 6/7/2011 6/21/2011 7/5/2011 7/19/2011 8/2/2011 8/16/2011 8/30/2011 9/13/2011 9/27/ /11/ /25/ /8/ /22/2011 Average hourly megawatts outcomes is the same: these offsetting bids do not add any net supply or demand to the day-ahead market, but can exacerbate real-time imbalance offset charges when hour-ahead prices diverge from real-time prices. Fig 8. Offsetting Virtual Bids Figure 4.5 Average hourly virtual imports offsetting virtual internal demand 3,500 3,000 2,500 2,000 1,500 1, Offsetting virtual bids by same participants Offsetting virtual bids by different participants Week beginning Offsetting virtual supply and demand bids at internal points

16 Price Spikes in Real-Time Market Spike No Spike All Hours 2009 count of hours 62 6,538 6,600 % of hours 0.94% 99.06% % DA/RT spread $ (349.06) $ 1.18 $ (2.21) 2010 count of hours 87 8,673 8,760 % of hours 0.99% 99.01% % DA/RT spread $ (343.68) $ (0.65) $ (4.06) 2011 count of hours 65 8,695 8,760 % of hours 0.74% 99.26% % DA/RT spread $ (341.80) $ 2.32 $ (0.23) 2012 count of hours 70 8,714 8,784 % of hours 0.80% 99.20% % DA/RT spread $ (365.55) $ 2.11 $ (0.82) 2013 count of hours 47 8,713 8,760 % of hours 0.54% 99.46% % DA/RT spread $ (335.67) $ 3.98 $ count of hours ,333 41,664 % of hours 0.79% 99.21% % DA/RT spread $ (347.81) $ 1.82 $ (0.97) Source: Data are from OASIS. Calculations are author's.

17 Fig 2. The Hourly Granularity of Day-Ahead Markets CAISO Ramp Rates: Example Load (MW) DA Schedule Assumed ramping 50,000 40,000 30, Hour 11

18 Fig 3. Real-Time Intra-Hour Load Ramp Load (MW) 50,000 40,000 30,000 Steep 5 min Interval Ramp DA Schedule Assumed ramping RT Load 15 5 min interval Hour

19 Fig 9. The Impotency of Virtual Bidding Load (MW) 50,000 45,000 40,000 DA Schedule Assumed ramping RT Load DA Schedule with Virtual Demand 30, Hour

20 The Loss from Virtual Bidding into RT Spikes Results in increased scheduling of generation Day-Ahead. Increases the DA price, which means improved convergence, on average. An apparent improvement. However, this turns out not to be welfare improving: 99% of the time this is a mistake as DA is already above RT, and extra supply scheduled increases DA cost. 1% of the time this is also a mistake because it is the wrong capacity. Price still spikes in the Real-Time market. More is not always better.

21 Summary spreads between the Day-Ahead and the Real-Time price will often arise due to the many necessary approximations differently employed in the Day-Ahead and Real-Time algorithms; while virtual bidders can profit off of these spreads, oftentimes they cannot help resolve the underlying problem; in these cases, profits earned by virtual bidders can be a purely parasitic drain on the system, adding to the costs paid by load; in addition, virtual bidders may add to system costs; convergence a narrowing DA/RT spread is an imperfect metric for evaluating system performance and the contribution of virtual bidders; virtuals may cause the average DA/RT spread to move closer to zero, and nevertheless all virtual profits are a purely parasitic drain, and, in addition, virtual trading has increased system costs.

22 Empirical Literature Overlooks This Problem CAISO Hourly DA/RT Spreads, Before and After Virtual Bidding AFTER DA/RT Spread, $/MWh BEFORE Hour of the Day

23 A General Problem Episodic. Disparate. Transient....Inherent.

24 Importing Financial Traders to Discipline Bidding the cane toad, Bufo marinus, was inten6onally introduced to Australia to control the introduced French's Cane Beetle and the Greyback Cane Beetle,[27] pests of sugar cane. 102 toads were obtained from Hawaii and bred in cap6vity to increase their numbers un6l they were released into the sugar cane fields of the tropic north in It was later discovered that the toads could not jump very high and so they could not eat the cane beetles which stayed up on the upper stalks of the cane plants. However the toad thrived by feeding on other insects and it soon spread very rapidly; it took over na6ve amphibian habitat and brought foreign disease to na6ve toads and frogs, drama6cally reducing their popula6ons. Also when it is threatened or handled, the cane toad releases poison from paro6d glands on its shoulders; na6ve Australian species such as goannas, 6ger snakes, dingos and northern quolls that aqempted to eat the toad were harmed or killed. excerpted from the Wikipedia entry on biological control