PhD Thesis: What I did, What I found, Why it Matters

Several years ago, I decided I wanted to write a Masters Thesis that didn't just sit on a shelf, collecting dust.  Call me idealistic, but I wanted to adopt a philosophy of "science off the shelf" 

[April 2016 Update: My PhD Thesis is now finalised! Download it here: Part 1  |  Part 2 ]
 


[download my submitted thesis - part 1]

[download my submitted thesis - part 2]

Back at the University of Oklahoma School of Meteorology, there was a running joke amongst graduate students that they could stick a $20 bill in their thesis, check back in 5-10 years, and find it still nestled between the pages, safe and sound.  While it was humorous, and mostly in jest, it reflected the mentality of postgraduate education/research work - what we were doing was mostly just to tick boxes & get our degree so that we could one day do something that mattered. In reality, our 'research' would sit on a library shelf, acquiring dust.

Yearly kWh production for a simulated 2kW Sanyo array in Oklahoma

Yearly kWh production for a simulated 2kW Sanyo array in Oklahoma

But I wasn't content to resign myself to this fate. I felt compelled to produce something that mattered, that would have real-world relevance.  So I brought on an industry partner, a local solar installer in Oklahoma, and put together a solar map for Oklahoma, not based on solar radiation - but on kWh hours of electricity generation, which immediately translates to payoff times and dollars. [download my M.S. Thesis]

It was a tremendously successful experience, and as I moved forward into my PhD project, I was very determined to not only maintain that approach, but to expand it.  Now, looking back, I can see that approach has paid off in great ways, which I'd like to share with you.  Because what I've discovered, what I've developed and what I am now positioned to do with the technology I created, is all highly relevant, commercially viable and ready to hit the ground running.

What I did and why

Australian PV installations by year, as provided by the Australian PV Institute

There is ~4GW of solar energy installed in Australia, which is dominated by ~1.4 million small-scale photovoltaic (PV) arrays.  These arrays are relatively small (average size 1.5kW in 2011, growing to 4.5kW by the end of 2014), and the vast majority of these installations are un-monitored.  This means that their minute-by-minute performance is not recorded.  In fact, the only information collected for most of these systems is the total quarterly production as reported on electricity bills.  Long story short, this lack of information limits the number of PV systems that can be installed in a given region (like here).

Measured PV power output, divided by that PV systems clear sky power output

Measured PV power output, divided by that PV systems clear sky power output

So what did I do about it? I developed a method for estimating the power output from many thousands of PV systems using a small selection of monitored PV systems (which report their minute-by-minute or hour-by-hour generation).  This method is called  “the clear-sky index for photovoltaics”, KPV. [read more]

Along the way, I had to do significant work in the field of solar radiation modelling, validating clear sky radiation models and developing a new type of solar radiation model fit to Australian radiation data (a "separation model").  I also had to show that my new method (using solar panels as a sensor network) was able to do the job just as well as professional grade radiation equipment.  

Pyranometer based methods versus KPV based methods.  The takeaway? Get the sites close enough (within 5km) and my new method is just as good as using a more sparse network of professional grade solar radiation sensors.  

It was also necessary to develop new quality control routines for the PV system power output data.  This type of data is messy, often provided by non-experts and hasn't been used in the way I've proposed before.  This was accomplished through the development of a new quality control routine called QCPV (now going through the review process in a major journal).

~200 solar PV systems installed in Canberra, which are reporting their data in real-time

Once I sorted out the quality control work, it became possible to work with data from many hundreds of solar PV systems, as well as scale my KPV method up to many thousands of systems.  So I forged ahead, using Canberra as a proof-of-concept, creating a city-wide distributed PV simulation of its 12,000+ embedded PV generators (based on December 2012 installation data).  

I then paired this simulation system with weather events that cause broad-scale, rapid changes in the power output of all of the PV systems at the same time [check it out].  It is these types of events which are the most likely to cause future grid stability problems.  The basic idea is that, when you have a wide-spread solar network, the negative effects of partly cloudy days are "smoothed out" by those systems being positioned over broad region, but during certain weather events, it is not possible to smooth out these impacts, because the cloud features are too widespread, sudden and thick.

"when you have a wide-spread solar network, the effects of partly cloudy days are "smoothed out" by those systems being positioned over large region, but during certain weather events, it is not possible to smooth out these impacts, because the cloud features are too widespread, sudden and thick"

What I did find?

Let me keep this as focused and brief as possible...

KPV estimates (color) versus measurements (black) under a positive ramp event

Firstly, I demonstrated that my newly proposed KPV method was much better than existing methods, showing that it was well-behaved under all cloud cover conditions, and performed well under positive and negative ramp events. [download publication][read blog post]

Next I found that for clear sky radiation models, operating in Australia, global clear sky simulations, are best computed by the Solis, Esra and REST2 approaches, while the Iqbal, Esra and REST2 methods are the most proficient clear sky beam models. [download publication][read blog post]

The Engerer 2 separation model at work (blue model estimates, grey observations)

 After that, I found that only the Perez separation model performed satisfactorily for high resolution (one minute) solar radiation data.  In response to this, I developed three new separation models, which gave slight improvements over the Perez model and greatly exceeded the performance of all other existing model techniques. [download publication]

Once that was handled, I compared radiation sensor based methods to my PV data based approaches, with a student project.  This study found that the approaches were equally as good for separation distances of 5km or less. Given that PV sensors are "cheap" (someone else pays for them) - this was a great finding. [download publication][read blog post]

Post QC KPV estimates, very tight correlation, great results

Then I dug into the development of the QCPV algorithm (quality control), demonstrating that the method I created can result in a 43% reduction in Mean Absolute Percent Error (MAPE) over the raw data. [pre-print coming soon]

Second to last, with another student project, we categorised the weather events that cause those large scale, collective changes in PV power output discussed earlier.  Positive collective ramp events (sudden clearing) were caused by Australian northwest cloud bands and radiation fog dissipation. Negative collective ramp (sudden cloud cover arrival) events were caused most frequently by the passage of cold fronts and thunderstorms.[download manuscript][read blog post]

Finally, I put it all together, with the city-wide PV simulation system, using it to simulate the changes in total power and energy output from these collective ramp events.  I was able, for the first time, to quantify (aka determine a representative number) the amount of power that (dis)appears on the electrical grid during these events.  For example, a thunderstorm event on 19 February 2014 removed 20.78 MW of power generation from the local grid over an 85 minute period, which equates to approximately 14.54 MWh of energy generation forgone over that period.  That's probably enough to change prices on the energy market - not very much, but as the solar installation numbers continue to grow, that influence will grow significantly.  

Here's the thunderstorm ramp event from 19 February 2014, along with some satellite imagery.

What is my overall conclusion?

The overall conclusion, is that the developed regional simulation system for distributed solar PV, made possible by an upscaling of my KPV methodology, represents a significant, unique and promising tool for scientific, engineering and operational purposes.  

In the simplest of terms: I built a very handy tool, with cheap inputs that can be run anywhere that solar PV systems are reporting their power output data.

Where next?

I have a full-time lecturer position ("professor" in the American use of the word) at The Australian National University, where I work in the Fenner School of Environment and Society (employed since July 2013).  I am using the freedom and security this position provides me with, to apply for funding to scale this simulation system up, Australia-wide.  I'll join it with the new Himawari 8/9 satellite data, and pair up with the energy market/utilities in Australia, in order to help large amounts of distributed solar to be added to the grid.  

You could say the future is sunny and bright (#punny).  And with HUGE amounts of solar being installed globally, the solar century is before us.  There are plenty of opportunities for this science to stay off of that dust shelf.  So I'd say, overall, this whole PhD thing has been a smashing success! Even if it was a bit of a wild ride.  I hope to have more exciting news soon - for now, I'll get back to working getting this simulator to run real-time in Canberra...

/*