Abstract: Testing a Real-Time Geothermal Power Contract Management Tool Performance in a Chaotic Wholesale Market
By Daniel M. Hamblin and Lance D. McKinzie
Overview
Our paper tests the resilience of real-time contract management software in a chaotic wholesale market for electricity. The software was designed for use with an innovative escalator price design to attract buyers, sellers, and brokers to electricity from renewable energy sources in the western continental U.S. An initially discounted price attracts buyers to the renewable resource perceived to be not as dependable as conventional capacity. The price escalates relative to the prevailing wholesale market price, to include a premium or margin over the contract term. Power brokering (with an escalator price) and renewable energy coupons (green tags) from an interested donor are tools for arbitrage. In a California market, one is likely to have a vendor or geothermal site developer with a steam-generatingplant that lasts 20 years or so, requiring an investment of ~$2600/kW in overnight cost (according toPlatts), and needing a payback of 10 years or less to break even in the out years. The escalator price compensates the seller with margin in the out years to help resolve the payback dilemma. Additionally, tinkering with theassignment and timing of green tags shortens the wait. Payback in net present value on a 15-year bilateral contract is from about 4 to 4½ years for a discount rate between 5½ and 7½ percent and green tags valued at $20/ MWH. W. Michael Warwick developed the escalator price concept that this and previous papers put to use. This is the third paper (with Hamblin as the principal author) about the software and rate design. The first described the software by its architecture and a specific application to flash geothermal steam generation sited in California’s Imperial Valley. The second used Latin Hypercube Sampling and Platts’ wholesale 7x24 prices to test the software’s vulnerability: to adverse selection of vendors or geothermal site developers and to moral hazards posed by dishonest brokers. The current inquiry raises the stakes for advocating the software and rate design by looking at how they handle wholesale prices higher, lower, and much less predictable than those experienced while the target market has been deregulated and subsequently re-regulated. Our inquiry also disclosed a weakness in the Sandia Latin Hypercube Sampling executable that we exploited by adding useful economic information to the samples and imposing statistical tests on the results.
Methods
The methods we used tailor inputs for testing the Escalator P software and process its output to ascertain how well it performed. (See figure.)
The adjusted Platts’ scenario prices, price deflators, and their paired correlations are inputs to the Sandia executable, in which we use restricted pairing proposed by R.L. Inman and W.J. Conover to make inputcorrelation as small as possible. The Sandia National Laboratories’ software, obtained under an exclusive use obligation, decremented our inputs, yielding prices under $100 in $0.10 increments, prices $100 and higher in $1 increments, and price deflators that were no longer unique for each month of data input. LHS Transpose converted 100 predicted outcomes for each of 240 months of price and price deflator inputs to 100 transposed 240-month samples and sorted each sample by decreasing value of the deflator. This formed baseline data input for Escalator P. In addition, we transformed the baseline data of each sample using decision rules that change the order of the monthly data in months with the same pricedeflator. (For the 180 sample-months used to test Escalator P’s vulnerability to principal-agent problems, the number of duplicate price-deflator observations ranged from 19% to 39% of the months, with an average of 31%.) Our decision rules created two alternatives to the baseline: Quadrants 1 and 2 in which the rate of inflation exceeds the rate of electricity price increase for duplicate observations, and Quadrants 3 and 4 in which electricity prices go up faster than the inflation rate in duplicate observations. Escalator P distinguishes 1 and 3 from 2 and 4 by the segmentation of magnet price vectors used by its clustering algorithm. Quadrants 1 and 3 use 193 vectors. Quadrants 2 and 4 use 97 vectors. LHS testing takes days of dedicated-computer simulation. 193 vectors take twice as long as 97. We wanted to know if the reward forincreased segmentation is significant. Our endeavor requires scrutiny of Escalator P frequency plots as they scroll down the computer screen and numerical results in files. We follow this by culling extraneous information from the files and using expert systems to diagnose what remains for the baseline and four quadrants. The expert systems decide which price scenario works best each month to minimize adverse selection and which is most tempting to a dishonest broker. They were extended from those of the second paper to accommodate the baseline/quadrant transformation and to yield statistical diagnostics for the duplicate observations alone in each sample. We produce summary diagnostics that contrast baseline versus quadrant results from the 100 expert systems each for the baseline and four quadrants. This and the extended expert system diagnostics support our paper’s final inquiry: whether or not a simulated real-time contract’s final settlement payment predicted for the four quadrants is statistically different from that for the baseline.
Results
Using the adjusted Base Platts’ scenario to compute the escalator price slate won a slim plurality as best for minimizing adverse selection, averaging 31% of the 180 contract months computed from 500 Latin Hypercube Samples. Similarly, the adjusted NYMEX was more often susceptible to moral hazard posed by dishonest brokers, winning in 34% of the contract months. While the Contract-For-Differences settlement payments over contract life were 8.8 times higher using the chaotic (adjusted Platts’) price regimes than they were using the pure Platts’ price regimes used in our previous study, the Final Settlement Payment at the end of the contract was just 2.1 times higher – telling us that the residual error in the Escalator P clustering algorithm increases about 1/4th as much as the monthly contract settlements that produce it. However, one transposed Latin Hypercube Sample predicted a negative Final Settlement Payment, an inducement to adverse selection because it requires an end-of-contract refund from seller to buyer. The refund is more than three times higher for Quadrants 2 and 4 than for Quadrants 1 and 3, an argument for the increased magnet-price-vector segmentation. Brokerage fraud proves to be a risky business under a chaotic price regime. While the average payoff for 500 MW worth of fraudulent brokerage ranges from about $109,000 to $111,000, a dishonest broker has a 16% chance of losing as much as $106,000 from the price slate common to Quadrants 1 and 2, and a 14% chance of losing as much as $79,000 from the price slate common to Quadrants 3 and 4. Finally, paired comparisons of the baseline versus quadrant-specific results yield 95% confidence that we cannot reject the hypothesis that baseline and Quadrant 1 and 3 Final Settlement Payments are statistically indistinguishable for 99 cases out of a 100. Paired comparisons of the baseline versus quadrant-specific results yield 95% confidence that we cannot reject the hypothesis that baseline and Quadrant 2 and 4 Final Settlement Payments are statistically indistinguishable for 93½ cases out of a 100.
Conclusions
We conclude that our Escalator P real-time bilateral contract management tool for flash or dry steam geothermal with an innovative escalator price design to attract stakeholders is indeed resilient in a chaotic electricity price environment. Its use could maintain geothermal growth in the renewable fuels mix while Enhanced Geothermal Systems move cautiously forward. We believe we could improve our software’s vulnerability to principal-agent problems through a “dominant-condition” gaming approach. Four dominant conditions occur to us: (1) Saudis/Venezuelans reduce production; (2) Hurricanes disrupt supply; (3) Tectonic/volcanic event changes outlook; (4) Political/economic downturn reduces demand. An expert system driven by the depth-first-search algorithm could weigh the probability of the dominant conditions and evaluate Escalator P resilience to principal-agent problems in the broader context of world and local events.