Message-ID: <27562051.1075856643857.JavaMail.evans@thyme> Date: Tue, 8 Aug 2000 02:26:00 -0700 (PDT) From: vince.kaminski@enron.com To: joseph.hrgovcic@enron.com Subject: Re: Cc: mike.roberts@enron.com, vince.kaminski@enron.com Mime-Version: 1.0 Content-Type: text/plain; charset=us-ascii Content-Transfer-Encoding: 7bit Bcc: mike.roberts@enron.com, vince.kaminski@enron.com X-From: Vince J Kaminski X-To: Joseph Hrgovcic X-cc: Mike A Roberts, Vince J Kaminski X-bcc: X-Folder: \Vincent_Kaminski_Jun2001_5\Notes Folders\Weather X-Origin: Kaminski-V X-FileName: vkamins.nsf Joe, I have written to Pal Quilkey to invite Christian Werner to Houston for a week to discuss how much of his research we can use. Vince Joseph Hrgovcic 08/01/2000 06:51 PM To: Vince J Kaminski/HOU/ECT@ECT, Mike A Roberts/HOU/ECT@ECT cc: Subject: Vince, I have inquired into Christian's climate models. It seems to me like a long term project. Chiristian says that his model might be able to do better than the Australian Met given that they ignore some of the variables he uses and that they are an "old boy's network" resistant to new developments, but I don't think the same can be said of most other weather services. As far as the NWS goes, they run their model on a Cray, they have several very talented PhD's working on it full-time, and even then, they can only get a dozen or so runs per night. In other words, it's a huge system. Replicating something even close to that will not be an easy task. That being said, I think there are related applications that we could look into. Since I've already promised the RAC group and the weather desk's PJM traders to put together a Vector AutoRegressive model of daily temperatures, I think it makes sense to see if something better than a VAR model can be put together, perhaps a very stripped down version of the kind of model Christian has. What I have in mind is something that would give us say, up to several dozens of "possible" temperature forecasts every morning, which would be calculated using actual climate models (as opposed to time series models). I would not use this as a forecasting tool (the NWS model results would make a far better "best guess"), but our ensemble could still be used to provide a distribution of temperature scenarios. This ensemble would have several uses: 1) we could price the book for the ensemble of runs and thereby obtain a more realistic daily V@R for the Weather Book (and eventually interface that with the power desk's V@R calculations) 2) we could use the ensemble of forecasts to relate temperature forecasts to week-ahead CDD and HDD distributions (for use in our demand swaps) 3) if Freese-Notis were to give us one of their qualitative forecasts, e.g., "expect the trough to recede" we could sort through the different Monte Carlo scenarions, find one in which the trough in question is receding, and use corresponding output statistics to download the actual temperatures corresponding to that scenario. Up till now, we haven't found a good way of getting a numerical fix on what different forecasts actually mean temperature-wise 4) we could use the associated visualization routines that come with such models to get animations of the evolution of historical weather patterns. The traders could use these in order to look for historical periods which roughly match current weather conditions and get an idea of what could happen; although these meteorological models would not be good for simulating month-ahead or season-ahead weather (trust me on this) we could still use the visualization technology to do to analogous seasonal animations. Just today, I've spoken with Dr. Jacobsen of Stanford University, who has written one of the more recent textbooks on climate forecasting , and who worked with UCLA's general circulation model. He says that getting a simplified version of MM5 (which is itself a simplified, limited-area version of the GCM models that NWS, COLA, and UCLA use) would take several months to implement, assuming the people involved are already well versed in the technology. Also, I have contacted AER, a Massachusetts- based weather consulting firm, and they tell me that they have a MM5 model running daily (only one run per night), and also some smaller PC models up and running. They are of course willing to work with us for a fee. Their MM5 version runs on a $200,000 parallel processor. I am open to your suggestions (or objections as the case may be). I'm not sure how the costs would be apportioned given that this would benefit all of Enron, and not just the weather desk. I am scheduled to go to Boston next week anyway, and would like to use the opportunity to visit with AER. I will of course coordinate any projects with Christian (if we get something like this up and running it will be more likely that he can get the computing power he needs to run his own Australian model). Joe