The Universality and Predictability of Technology Diffusion

Show video

I'm motivated to study this because of uh climate  change and actually this all started back in 2009   when Dan arisu who was the head of the national  renewable energy laboratory and who also is a new   Mexican um came here with a group of people from  the national renewable energy lab and said hey can   you guys help us think out of the box and after  a few hours of talking with them it became clear   that the whole problem of adapting to climate  change hinges around technology which Technologies   are going to cost what when and so ever since  then I've been working on this uh first with   a group of people here some of you a few of you  may remember ban NJ uh who by the way disappeared   nobody knows where he is now um long interesting  story um but okay so this is just the de ongoing   development of that work now it's it's clear that  we need better guidance in the energy transition   uh I would argue the track record so far is  poor um standard integrated assessment models   which are the main Workhorse tool people use for  this compute an optimal transition path based on   a given climate emissions Target so that's the  question they're asking and um I think it's a   peculiar question to ask because it really only  makes sense if we had a global dictator who could   say this is what we're supposed to do and um if  you believed these arguments about that you know   we we have to make a tradeoff between well-being  and transition that might make some sense but um   people don't do that and so models are untestable  you poke at these guys and say look your model   hasn't been doing very well they say oh well it  it it's not designed to answer that question and   just to illustrate here's a picture of projections  that have been made um uh so so here is the actual   cost with the black dots of um photovoltaic  solar energy um these red and blue lines are   predictions made by various models of that type  the the black line there is the trend line from   1980 to 2020 and what you can see is like here's  a forecast made in 2000 for 2020 and they were off   by about a factor of 50 okay so they were off by  a factor of 50 and they were too high on the cost   and they've been making forecast since even before  that and they've always been too high way high and   they're high again and again and again and again  and I used to make the joke that the people at   the International Energy agency somehow they  can't plot data on log lmic scale because when   you plot the data on logarithmic scale the trends  are apparent these by the way are the projections   they made for improvement rates that were reported  in 2014 but mostly around 20 thou uh 2010 for how   much were going were were was solar energy going  to drop in cost over the next decade and you can   see there were about the order of 3,000 different  projections none of them came close to the actual   value which is that black line now why are these  models so off well to do this I already said   they're asking the wrong question but they also  have to make ad hoc assumption so these models   can get quite complicated I mean the remine model  at pick probably I would guess thousand person   years have gone into building it um and they have  to assume things or they assume things when they   make them like they they put in solar photovoltaic  and they use a learning curve I'll describe what   that's about in a minute and they um uh but then  the model says oh do nothing but invest in solar   energy well oh that can't be right prices can't  get that low so they put in floor costs so these   green lines are examples of floor cost in various  models of this type the black uh line again is   PD system cost you can see it just repeatedly  punching right through the floor costs that they   assume were impossible to achieve um okay now so  our model modeling philosophy is different we we   want to ask answer a different question we want  to try to predict what will actually happen and   how can we nudge it and I don't mean nudge in  the sense of psychology but what can we do to   incrementally make things better we want to make  testable falsify predictions one of the problems   with these other models they're completely  unvalidated they just make projections they   don't call them predictions because they don't  want to be too assertive and so we test everything   out of sample we try and understand how good the  predictions are and we want to build reliable   models and we're actually going to do things much  more simply now those of you who were at my talk   last year or who were at my talk last week will  have seen this picture doesn't hurt I thought I'd   put it up again because some of you weren't  at either of those and um and so this is the   history of the global Energy System over the last  140 years and on the left we see the costs on the   right we see deployment and we see the different  kinds of Technologies so say that Dash black line   up there is the cost of Coal Energy um the no  sorry the brown one is Coal Energy the black   one is oil the gray one is gas you can see that  those they wiggle around a lot but there's no   obvious TR Trend um in contrast some recent uh  contenders like solar PV the orange one um uh   hydrogen p2x fuel meaning hydrogen-based fuels the  green one um wind that blue and there they're all   coming down in cost in a fairly regular way  in fact solar photovoltaic energy costs one   10,000th now of what it did in 1958 when it was  first de loyed in the Vanguard satellite um and   then over on this side you see uh the deployment  of these energy sources that green line across   the top is traditional biomass the dash Brown  Line is coal the black and gray lines are oil   and gas then the Blue Line Is hydro the pink line  is nuclear the dark green line is biopower blue is   wind electricity purple is battery and orange is  solar solar PV and then the green one is um again   p2x fuels so again something remarkably different  is happening you see that historically coal and   gas came up exponentially but at a at a slow rate  compared with what solar and wind and batteries   have been doing um it's also notable that nuclear  energy initially shot up at a fast rate and then   plateaued out um okay so that's just background  and and in a sense what I'm talking about in   this talk is projecting these kind of curves so  how do we project the cost curves and how do we   project the deployment Curves in fact I'm not  going to just say project I'm going to say how   do we predict the future of these and how do we  do it in a probabilistic manner so we understand   how good our predictions actually are Jim what's  useful useful energy ah so that's a it's a bit of   a tricky thing because you have to make some  assumptions but if you just ask about what's   usually called final energy that's how much energy  is there if you could actually get every Jewel of   energy that's in it to do something useful useful  energy takes into account typical conversions so   if you're running a um uh normal you know gas  fired or sorry uh internal combustion engine   then the efficiency is about 25% so you take the  final energy in the gas and you divide it by four   sort of like a free energy it's it's kind of like  free energy but it's a little little rougher than   that because it really depends on the way it's  being used but we did that even though there's   some ambiguity in how you do it because we wanted  to have an Apples to Apples comparison people   there's a lot of confusing crap in the literature  where people uh talk about final energy which is   kind of irrelevant now I'm going to so somewhat  zip through the first part of this talk talking   about cost because I've been working on that for  de decades and some some of you at least have   heard it before so but because I don't want to get  to the the topic of the title I gave but I still   want to say something about it because I think  it'll interest those of you who haven't seen it   um so first of all in predicting it's worth noting  that techn though you might think oh technological   progress impossible to predict because it's  Innovation innovation's unpredictable by its   very nature yes but the rates of progress are  highly predictable as Gordon Moore showed we'll   get back to that but just to illustrate here's  the cost of various things over 20-year period in   the United States ranging from hospital Services  which you know over that 20-year period went up   by more than a factor of two to televisions which  at least if you adjust on a quality control basis   dropped by uh 95% in other words they went down  by a factor of 20 in cost huge differences in how   these different things behave also notice they're  very persistent televisions didn't just suddenly   drop they drop every year um and we're back to  Moore's Law which is quite a remarkable thing you   know and originally Moore made the statement  that the density of transistors was going to   double every 18 months he later on adjusted it say  every two years but since then that prediction has   been remarkably good though we've always known it  was going to hit a wall around now when um things   actually reach the quantum scale so it's unclear  what the future of this law is uh it seems like   it's got to stop um and so this just shows you  what I said that even if the Innovations are   unpredictable and actually I remember being in a  conference here David was in the room um Doug was   in the room and it was on Innovation and invention  and so forth and there was a guy standing up here   um who was an actual chip designer and he said  you know you guys you all think Mo's law it's like   god-given it's just going to happen every year  and I can tell you every year we we'd feel like   we were facing a brick wall and then we'd pound  on it and eventually we'd figure out a way to   solve the problems so so well Mor's law you know  some people say it's a self-reinforcing prophecy   I don't believe that first of all it obviously  wasn't before Moore wrote it down why did he   write it down because it was happening before he  noted it and and secondly well I just think it's   caused by other things um though it is certainly  extremely useful and this was driven to me home to   me by Alvi Ray Smith my fellow New Mexican who  is you know one of the brains underlying Pixar   and he said at Pixar they had all the technology  ready about 5 years uh before they could actually   make a movie and they said look we just have to  wait we'll we'll polish it up and make it better   but we just need the machines to run faster and  as soon as that happened they made Toy Story so   so they were able to plan in other words around  that and in fact Mo's law did mean that the rest   of the industry could plan assuming that CPUs were  going to be faster in a certain date etc etc it's   extraordinarily useful to have a law like this  now uh there's also rights law which is actually   much older it it's a little bit different it says  what Wright said observing he was in the aviation   business he said that any given airplane Factory  every time its cumulative production doubles the   cost of producing the airplane drops by 20% so  works good for airplanes works good for tons of   other Technologies it also works on the global  scale as well as on the level of individual   factories um so different law and it's it's it's  now a power law which actually means that the rate   of progress is slowing down in the sense that you  have to have exp it's every exponential increase   gives you a 20% drop okay now so we extended  this we developed a probabilistic method for   forecasting technology costs based on historical  data we did a preliminary version of this here at   SFI with baa and and then we extended it to do it  better better um with franu Leon and myself and   then some other collaborators and we tested the  method by making 6,000 forecasts for 50 different   Technologies it worked pretty well and we applied  it to three scenarios for the green transition so   again I'm going to pace through this a little bit  fast you can always stop me and ask questions but   our forecasts look like this so notice we're not  making a point forecast And since this forecast   is kind of trivial you can see here's past data  for solar PV starting in 1980 um you can see it's   Wiggly but we draw a line through it we actually  just go from the first point to the last point   and from that last line our model says well these  are the probabilities of the different outcomes   where there's a 5% probability of being out in the  white zone and each of those bands correspond to   a 5% probability so it's a probabilistic method  and we actually tested the probabilistic method   right so yeah in other words we believe we have a  pretty good idea of how good the forecasts are and   to illustrate we actually wrote the paper where  that last circle is in 2013 that was at least the   most recent data and you can see the subsequent  data points and they're you know well within the   band that we predicted they should be within um  uh you know we have here we we had we then used   this method to look at actually 14 different  Technologies IES some of the ones like fossil   fuels we just used something simpler because they  don't follow this rule um the technologies that   do follow these kind of laws are somewhat special  but anyway we we built some scenarios to ask about   what's the cost of the climate transition be and  is there a way to get there that's cheap so we   looked at the key Tex of solar PV when batteries  p2x fuels we extrapolated the deployment Trends   we ass we phased out fossil fuels over the next  25 years that's an assumption I'll come back to   it in a minute and we used p2x fuels for energy  storage so we stored enough we assumed the World   by the end of this 25-year period was going  to have enough stored hydrogen that if the sun   completely stopped shining so it it's just night  all the time no wind blowing we could run the   planet for a month on the stored energy not clear  exactly how much you need maybe you need even more   than that because of inter but it's still a  pretty fair amount we also dealt with all the   hard to decarbonized sectors like heat shipping  Air transport and so on and we did all those   with p2x fuels in other words you make um you use  solar energy or wind to make electricity to make   hydrogen to make fuels and then you burn those  fuels for those applications and we tried to make   very conservative cost estimates because we didn't  want anybody to accuse us of you know boosting our   favorite techs so here are the three different  transitions that we looked at uh your labels are   gone but there's the fast transition here the slow  transition in the middle and no transition on the   right or slow trans very slow transition you can  see we enormously ramp up electricity which is   shown down here and that orange band is the role  solar plays the blue band is the role wind plays   the green band is the role hydrogen p2x fuels  play and all the other stuff is you can see the   fossil fuel is getting phased out in that plot in  the upper right upper left corner so basically we   assume in about 20 to 25 years these take over  now is that plausible in our paper we showed   these three plots you see solar wind batteries  and p2x electrolyzers we drew a red trend line   it's plotted on semi- lar scale you can see solar  so far and actually if you just extrapolated it   exponentially it take over quite a bit faster than  we assumed we assumed in that upper blue line that   it starts to taper off right now but but and then  it tapers off even slower uh no even faster in the   slow transition or the no transition um okay and  this is showing um the re the top panel the white   space is nonfossil fuel and so you can see by 2040  non-fossil fuels are largely taking over the blue   one is the is the fast transition so in the fast  transition most most energy is being provide not   being provided by fossil fuels by 2040 these these  predictions are all probabilistic so here you see   the fast transition slow transition no transition  so there's still quite a bit of overlap in there   I think we've been cautious in making sure our a  bands were big enough uh that we're going to stay   within them and this shows you the relative costs  over here so if you say take a 2 and a half% um uh   discount rate then we save the order of 10 to 12  trillion doll so we save a lot of money in other   words by doing this why because since we go down  these learning curves since we Renewables let's   just go back to that earlier plot if these Trends  continue you see solar in particular wind at a   slower rate by 2040 or 50 they're much cheaper  they make energy much cheaper than it's ever been   and because of that big savings we end up actually  saving money by making a rapid transition now I   want to point out that um oh and wait let me show  one other thing the plot over on the right here   shows greenhouse gas emissions due to energy  generation and you can see by 2060 they're at   zero they're actually pretty close by 2040 um  now I just want to beat my chest a little bit   in 2010 I published a forast in nature that um the  cost of solar PV would be cheaper than cold fired   electricity and um I quote The Economist from 2014  four years later solar power is by far the most   expensive they thought solar energy enthusiasts  were nuts and yet my prediction was right although   people it really irritates me again and again they  say nobody could have predicted this and I go all   you had to do is look at the data I didn't have to  be a genius to make this prediction I just had to   look at the data and believe the data which nobody  else seems to have done or few others did but but   not the main people in the IP CC and so forth and  by the way this just shows the track record of the   simple method we're using these bands in each  plot are the 90% quantiles so you can see and   making forecasts at different points in time for  solar wind batteries p2x electrolyzers you can see   the forecasts for p2x electrolyzers are still  kind of shaky based on that data alone that's   because the technology has not been deployed  that much yet on the other hand that technology   is a lot like a lot of other technologies that we  have lots of data on and and so it's the behavior   we're predicting is is in line with that namely it  should improve at a fairly decent rate um so you   can see these forecasts all did pretty good for  all four Technologies and okay I'm going to skip   this slide now I'm doing pretty good with time um  so what about diffusion everything I was talking   about until now was about cost and it was under  the assumption that we we the diffusion that is   the deployment of these Technologies happens at  a certain rate and then we got cost predictions   depending on the rate at which that happens but to  deal with climate change we really have to do this   pretty fast so you might immediately ask which of  those three scenarios that I showed you is most   realistic so you know it's I suddenly realized  a couple of years ago that nobody has done this   problem right um so question is can we predict how  quickly these will diffuse thus how quickly will   the transition happen well let's just look at the  historical data a little more closely so here we   see solar PV and wind we see the deployment since  actually 1976 for solar 1984 or so for wind and   semilog scale we draw a line from first point  to the last point for reference or or uh linear   regression and you can see that well it's kind  of Wiggly but roughly speaking these have both   increased exponentially uh for the last 45 years  and uh and at a pretty hefty rate the order of 35%   per solar and a little less than that per wind  and if you s at the curves there's maybe some   hint that they're starting to Plateau out but  it's not clear and notice solar in that period   of time it's increased in in its deployment level  by a factor of a million six orders of magnitude   so huge and and by the way since we're on this  slide this is an important thing to keep in mind   people go oh well you know there's still not that  much solar and wind it's only a few per of global   energy generation but exponentials with high  rates of growth are tricky it's easy to let   them fool you because they're small and they're  small and they're small and then suddenly they're   really big so these are about to appear on the  stage in a major way but you go well how do we   know they're not just going to flatten out it's  totally possible okay well so to do that we start   getting into technological diffusion Fusion if you  look in the literature there's a long literature   saying the deployment follows s-curves and so in  other words there's some kind of s-shaped curve   for example the logistic function that I plot  down below where in this function you have that   the deployment at some point in time is equal to  the ASM totic deployment level Time 1 plus expon   exponential of minus K which is a rate constant  time T minus t 0 t 0 is a location parameter that   is where the S curve is and so you can take data  and fit these kind of curves to the data and and   there's a whole kind of story incubation rapid  growth maturity blah blah blah um now people have   tried to use S curves to forecast what solar  and wind are going to do and here I give you   some examples uh let's maybe start on the top left  panel so there's four different studies that are   colored in the dots are the actual deployment  levels for uh electricity Generation by solar   energy um those gray lines are fits that were  made at different end dates so in other words   we go along and pretend the data is you know  we're stopping in 2015 we make a Fit Stop in   2016 we make a fit so we make fits every year  just the only thing that changes we add one more   data point every time we do the fit and you can  see they're all over the place so obviously this   doesn't doesn't work very well um but you'll also  notice the four published studies that I'm going   to beat up on a little bit here are all well below  that curve I'll come back to that in a minute and   you know some of these made big claims these guys  Kramer and Hugh hey I spoke at a conference at   Oxford in 2014 maybe and I was saying you know  look we're solar energy is going to get really   cheap etc etc they said oh you're completely  full of [ __ ] uh it's just can't diffuse it   can't take over more than 1% of the Energy System  why because of blah blah blah blah blah about you   know there's been no example that did that I and  that got published in nature um so all right then   here's a couple more studies um similar kind of  story and there's wind wind's a little different   a few of them actually are on the high side but  okay so it's easy to to um generate foolishness   by just using Lea squares fitting on this this  data and here's another example that that was at   National level Global level data this is National  level data where things get worse there we see   Austria where we see two fits one made in 2019  one made in 2012 you can see they're both utterly   different and you can see the data starts to get a  lot bumpier and in fact if you you know delve into   it you can ask what happened there well um Austria  had a generous policy here favoring the deployment   of um this one is wind and then they remov the  policy and then they instituted another policy   and so you can see these curves get bumpy at the  national level somehow though when you average a   lot of countries together it kind of Smooths out  and just another example of the garbage that's   been published um this is a paper by chirp at all  um and they're claiming that when you look at the   historical data doing a lot of mumbo jumbo that um  you can see that it's just utterly impossible that   solar and wind are ever going to provide the  energy because they had actually a good idea   the idea was let's look at the leaders the early  adopters let's see what how fast did they adopt   and what level did they Plateau out and let's  assume the others are going to be kind of the same   and then we can extrapolate from those leaders  not a crazy idea except the difference between   the leaders and the laggers is not big enough  to really tell the difference very accurately   the data is super noisy and you easily get absurd  fits like they say oh they include only the cases   where things have reached maturity and whether or  not something's reached maturity like changes Year   bye it'll first reach maturity indicating it's  achieved more than 50% of its deployment level   and then next you oh nope wrong didn't achieve 50%  and back and forth all right and we actually have   now written a paper criticizing all this stuff  I hate writing papers like this because you know   it's not pleasant the people that you criticize  get extremely angry but we kind of had to do it   because back to Moore's Law it's planning for the  pace of the transition if you believe these guys   then we shouldn't be investing too much in solar  and wind because they can't do the job they're not   going to get there quick enough etc etc ET so it's  it's a sort of high stakes thing and as I urge   them we should leave our egos behind unfortunately  that hasn't happened um but but so we go through   and and and just show unless you're very careful  to to not do overfitting data selection bias and   make other AD up assumptions you can lead to you  can get unreliable conclusions and I just wanted   to make it sure that it was clear that back here I  mean notice these predictions have been completely   falsified because they predicted the ASM totic  level was here and it's already here um okay   now this is a hard problem and there and there's a  good reason why nobody's done it right yet several   good reasons first of all the data is terrible  it's hard to come by a few you know fantastic   Pioneers have gathered some data and we've made  heavy use of them but it's still the data sets   are pretty bad the time series are also short um  you can see this let me actually go back and make   this point here you want to forecast these curves  based on this data this data goes not quite but   almost to present I guess it goes up to 2021 but  the thing is in an s-curve notice there's not a   lot of bending in an s-curve things start to bend  away from exponential growth but because well let   me just say this this function in the limit  where T is small it's just pure exponential   growth because this term is negligible right it's  only later that that this term becomes comparable   to this term and that's when things slow down but  just imagine trying to predict where this is going   to deviate it's pretty hard to do that okay now  the growth is inherently non-stationary when we're   in costs we assume there's some process that's  like Mo's law or a generalization of it or rights   law or a generalization of it there're stationary  Moors law in particular but growth is inherently   non-stationary because well you know there's  uh you know childhood middle age and maturity   um the deviations from exponential growth are  small in early stages I just demonstrated that   and as I'll demon demonstrate a minute the noise  is heterosis meaning the size of the amplitude of   the noise changes during the life cycle of the  technology the noise is heavily autocorrelated   meaning if you're above the standard S curve you  probably will stay there for a while the the noise   fluctuations if you get a big one the next one's  probably going to be big two and the fitting is   strongly downwardly biased that's why we saw  most of those forecasts being on the low side   now so how do you think about this um you know I  remember I was making a presentation for trying   to get funding to think about this stuff and and  one of the people I I said well you know um he   said well we know why solar energy got cheaper  and I go we do and he said yeah it's because of   the Chinese and I go well yeah but it was getting  cheaper at the same rate before the Chinese even   thought about making solar energy you go but so  I find over and over again people have a there's   always a story with solar energy okay we used  it satellites poles then you know a bigger Niche   because people started putting little widgets  on their lawn and and uh blah blah blah you know   the Germans put in feed in tariffs and then the  Chinese stepped in so that's the theory I would   call one damn thing after another in contrast  the logistic equation which I suggested should   fit these curves um if we were talking about say  bacterial growth we could write down exactly what   the bacterial doing how they reproduce the rate  at which they die we can solve those equations   which are roughly you get a differential equation  like this there's a little death term too and we   can come back to the modifications maybe in the  questions and so that's a deterministic process   and once you've observed the bacteria in the  petri dish long enough you can predict what their   future population's going to be fairly well now  I do want to emphasize this work as preliminary so let's begin by making a model based on  the logistic function we took the logistic   function and we've done one thing we've added  this term over here which is the noise term and   the noise process is now an autor regressive of  order one process meaning if if that makes the   noise persistent and this parameter row adjusts  how persistent the noise is if row is one it's   so persistent it's like a random walk and if  row is close to zero then it just reverts back   really fast and we find by the way it's quite  persistent row should be like 08 um so and and   then this noise here is just a gaussian IID noise  so that's the model we're going to investigate and   then let me see what the next slide is and then  we can do a trick that physicists love of making   a collapse plot so we to we gathered data on 47  Technologies when I say we I mean my graduate   student Ben vagen vort and um uh with some help  from Brendan Tena and Leonard uh bartner um and   and all the people who collected the data that  they gathered originally but we had gathered from   a lot of sources so we have 47 Technologies now  and we plot them in non-dimensional coordinates   so let me go back and explain that so  non-dimensional coordinates you see um   K the parameter K has dimensions of one over  time as it must because it's multiplied by   something that has dimensions of time similarly  L has dimensions of number of things so we plot   them in units where L is by definition one p is  all time is measured in units of K so if things   are slow then we we compress time if things are  fast we expand time so that everything's on the   same scale and we set we locate them all in the  same spot and so when you do that you see you   get an interesting collapse plot this is plotted  in what's called a Fisher pry transform meaning   we take this sigmoidal scurve and you make a  transform that flattens it out so now in in   Fisher pry coordinates the growth is linear um  and and so you can see that over in this part   you see what looks like a plausible straight line  but in this part you see all the Technologies are   looking pretty different so when we first did  this we said aha because the literature talks   about stage one stage two and stage three  technological growth roughly you know early   middle-age maturity and so there we have a way  of determining where stage one is because we can   detect roughly where this transition from  this really fast Behavior at the beginning   often in a you know sort of preul commercial  stage and then when it gets on an remember this   is a fish in Fisher pric this is roughly speaking  exponential growth and and and then flattening   out into maturity so anyway thought great now we  can isolate where stage one and Stage gives way   to stage two for all these Technologies but then  we realized that there's another hypothesis oh we   can make which is called a GT's growth process  so in a GT's growth process the derivative is   some rate constant K times the log of LX * X in  contrast remember the logistic one I did before   is looks like this now you know so somehow it  seems like there's at least some suggestion   that Technologies don't this is maybe not quite  as good is a scom process in bacteria it's pretty   easy to understand why this is true the KX part  is just because that's exponential growth and   then the one minus X is that as the bacteria fill  up the petri dish they have less and less food   so they can't grow exponentially anymore but  so this is some hint that the gumert process   may be better um by the way these are just to  illustrate four energy related Technologies   and maybe I should mention in this set we have we  have all kinds of stuff uh some of them you might   even question whether we should call technology so  we've got canals railroads internet all that kind   of stuff washing machines cars nuclear weapons  ammonia synthesis we even have um nuclear missiles   and monasteries so um they all seem to follow  very similar S curves now one thing you realize   when you when you once you make this collapse is  aha there's some universality and in fact back to   my question of how much is just one damn thing  after another and how much is the terministic   growth well we kind of see it here the signal is  that black dash line the noise is the variation in   all those gray curves and the answer is that  the bulk of the variance is explained by the   deterministic function this is true even though  you know because back to Sol and when people go oh   but you know there's all this we have to have all  these policies in place we have to do this we have   to do that and it really depends completely on the  policy how can you even make a forecast without   that well all these other Technologies you look  at railroads start reading about the history of   railroads there were huge battles over rights of  way very political thing um uh a lot of corruption   etc etc etc but so so it's partly signal and  partly noise but the signal Act is more powerful   than the noise now okay let's put this and make  forecast and um so here's for solar PV now first   let's ask the question what this suppose you have  a strong PRI about where it's going to land you   for whatever reason believe that solar is going to  land let's say the current um uh the current level   production which let's see this would be a the  power sector in 2022 so that's where it's going to land that's all right no wonder yeah much better  um that um okay suppose you believe that's where   it's going to land then you can ask what's the  probability for different outcomes given and here   we're assuming a logistic function and well look  up at so the our prediction would be we'll get to   80% of that ASM toote in 2037 in other words 13  years from now solar energy will be generating   80% of the power in the grid and those bands there  show you one and two standard deviation variations   so the prediction is not perfect but you're at a  95% downward prediction would be 2036 and upward   there would be 2030 2040 right so surprisingly  small band why you start plotting these curves   against each other you see they're not that  bendable we actually tried applying functions with   more than three parameters and by the way already  with three parameters it's extremely difficult to   get a fit but but you know you don't actually need  those you get just as good a fit with these simple   three parameters because once it's in motion  there's a lot of inertia in this deterministic   growth laws um now the plot on the right shows  the preliminary forecast for solar PV capacity   without assuming a prior and this is I emphasize  preliminary again because our forecasts are going   to change because we're tweaking up the method  and I could come back and say a lot about all   the work we've had to do to deal with all those  problems um but there's our preliminary forecast   and we're actually forecasting that it's going  to land more over in here and um so you can see   the plot for yourselves yeah so that ASM tootes  around 16 terawatts um now I want to leave room   for questions I want to emphasize I'm not saying  policy doesn't matter in this is all just going to   happen inexorably policy does matter just look at  the Wiggles in these curves a lot of those Wiggles   were probably policy driven um so policy matters  um on the other hand the good news is we've got a   lot of inertia going uh we do need to deal with  things like grid expansion fossil fuel politics   land use is a big one because I'm leaving out  the 25% of emissions that comes from agriculture   mostly cement but I I still think we're to see  energy get cheap and it's going to happen fast   um okay it's not just that okay we uh it's cheap  low volatility of energy prices you could see the   way those prices bump up and down it's actually  very bad for the economy when Energy prices are   fluctuating all the time you can't anticipate what  they're going to be a lot of things you want to do   depend on what energy prices are going to be and  so we should have much less volatile Energy prices   for reasons I can explain if you want better  energy security for anybody that has sun or   wind which is almost everybody low pollution uh  much lower pollution lower environmental impact   in general and much more sustainable and most  importantly no greenhouse gas emissions so now   you might say oh but you're just extrapolating  time series that's true that's all I've done   we're building models that do other things I  think I already showed this slide in my talk   last week um we by the way looking at the labor  transition this is showing that we actually have   a big the good news about the transition is that  we're going to get a lot more jobs it's actually   a job increasing thing at least for the next two  decades the bad news is it's a job bubble because   a lot of Renewables depend on construction and  construction you do the construction and then   you're done and so it's it's going to continue to  grow at 20 or 30% per year until it plateaus out   and energy growth only grows at about 2% per year  though it may start growing a little faster once   we get there because it's going to be so cheap um  so we're building just as a you know a trailer for   future research uh agent-based mod of the energy  and power sectors the agents are real energy and   power firms we have a massive data set showing for  the last 25 years 90% of the power plants and oil   fields and everything in the world together with  who owns them and and data about costs balance   sheets of companies so we're putting that together  to build an agent-based model where the agents are   the energy and power firms every year they may or  may not invest in new assets like they might drill   a new oil field or they might build a solar farm  and then we simulate the market every year because   the landscape changes we simulate the market we  find prices and quantities sold and we estimate   the profits of each of the companies and we  predict the future cost of Technologies we do it   year by year and now we have a policy laboratory  because we're going to do this for the whole globe   we just got a $4.6 million NSF Grant Hallelujah  and um and so we're going to do it for the whole   globe for um uh you know solar wind oil gas coal  and hydro so that we can really look at the entire   Energy System how the pieces compete with each  other and put in policies and how see how the   system responds to those policies yeah yeah um  don't so I'm a little bit confused your message   before a little bit was that policy doesn't matter  look at no these things we only care about the   Technologies we could just do the Plus without  paying attention to all the corruption around   railroads etc etc but now you're saying that um  actually the main motivator is to figure out which   policy to use yeah so I didn't say policy doesn't  matter I'm saying that a lot of the action is just   in the technology itself but policy does matter  because as I said there's a lot of fluctuations   around these curves and and you could see actually  quite graphically on the Austria one I showed you   back here look at Austria to just to see that  policy matters again there's a favorable policy   here the policy gets shut off there policy gets  reimplemented here so policies do matter it's just   that and and all this data included lots of policy  stuff in the past some of it more than others so   it matters but it's not it isn't everything  that's my point okay so question that and   then I'll shut up um you said you are not going  to assume that we have a dictator for how the   world right things bummer um but uh then you also  said that those policies at the level of Austria   for example that when you then look at the  entire world all that washes out so who would be   implementing the policy that you are investigating  yeah so every country has their policy actually in   the United States frequently states have policies  even cities can have policies and so we're going   to do this as best we can given how complicated  the policy landscape is but we want to build   this model for several reasons one is because I  want to demonstrate to the fossil fuel companies   that they're about to go out of business unless  they get on board quickly uh I think that will   help accelerate the transition um uh but but but  in general we want a policy laboratory where we   can play with it so we could say let's assume  Germany has this policy the US has this policy   China has this policy what happens and we can  now play around with policy scenarios that may   be quite complicated um so hopefully that answers  your question and so anyway this is the way this   model works we we um basic elements are the firms  and the plants we you know have an investment   module The Firm some firms go out of business we  update the costs we run the energy Market we go   around the loop and we're already doing this for  Texas power grid um we picked Texas because Texas   urot as it's called you can see there's a little  bit of Oklahoma in there and you're missing part   of West Texas but uh it's its own power Universe  it's an isolated power grid so it's we can model   it and and we can we've seen that we already have  a pretty decent model we can predict things about   the generation mix um we can predict things about  energy prices the predictions aren't perfect yet   but um they're I I think it's all quite doable  uh now I think that's the end of my talk thank you yes question is in your proposed model is the  policy endogenous or is imposed in our proposed   model you mean the one I was talking about at the  end yeah it it's it's imposed the policies are at   this point exogenous now my student Penny m a  student I mean my student as of eight years ago   Penny M who's now at the World Bank has um been  um studying policies and I think there is some   hope for somewhat endogenizing policies because  what she did is get a big database of policies   and classify the policies according to types and  look at the time history of the policies so that   she can actually understand which policies are  feasible for which places when and so with that   coupled to this we may be able to really uh go  even deeper sorry I'm just any other questions   Andrew do you have some intuition for the gobert's  growth process where you have now the difference   of the lws versus the actual difference between  the so you have in the gos gr the logs right   yeah logistic you just have the difference of the  two yeah do you have a good like analogy for the   intuition behind that of what no I I would love  to have some um you know I um I it's it's a good   question and I I don't I just we're thinking  about it quite hard but it's a great question   Sam thanks for great talk down um uh in uh you  mentioned canals and monasteries yeah uh and is   it I mean but you also described the inertia and  so on as a technological inertia well clearly the   case of canals is not primarily you're talking  probably is it canals in us actually globally   globally yeah that that's driven by public policy  uh and and and whatever monasteries are driven by   it's not technology so yeah is it just a quirk  that that they also work well do that suggest   there's something else going on that's making the  S curve happen that shouldn't be called technology   yeah probably the latter I think there is some  Universal process back there um and by the way   we might have a hard time distinguishing logistic  from gobert's technology by technology because the   data is pretty noisy but but you know on one hand  yeah canals are heavily driven by policy but on   the other hand there's a period where people went  wow canals they started digging a lot of them at   least in the developed countries and and um and  then actually you can see that as canals are   before they fully reached maturity railroads come  in and out compete canals a lot of the legislative   and you know competition was between canals and  railroads um so and and of course monasteries   again I I I don't know it's now we're talking  about cultural diffusion and I think it's just   that when you have growth processes with a limit  and where where there's a possibility of some   kind of replication process you know you make one  Monastery and people go wow I want to join a Monas   too and they go off and start another one  so somehow it's infectious in some sense so   I think some of the qualitative elements are  probably very similar I just Sam I think it's   David talking about confusion not the invention  yeah this is diffusion just a fusion and I think   for invention your criticism would hold right it  wasn't even a criticism not a the observation yes   cares monies canals integrated circuits that's  a common mechanism that's a that's a mechanism   of persuasion WRAL inertia here because you  have the same yeah whether it's a monastery   or whether it's a sh photov voltaics then clearly  that analogy doesn't hold that I mean that's not   exactly a criticism I'm I'm using technology  in a very loose way that's that's I should have   made that clear so I would call monasteries  are in some sense a social techn technology   um nuclear missiles are te okay that's obviously  technology but canals are a technology sentences   you said about people catching on and so on  yeah that's a social technology yeah there's   something similar going on yeah yeah yeah call  technology but whatever it is I yeah we have we   need some nice name for these things whatever  they are that seem to behave the same way it's   incredible that found that out Melanie so you  should like the price of nuclear is prettyat yeah I'm wondering if that's more a policy  issue or a technological issue and do you   do you think that nuclear has any place  in this yeah yeah so first of all it's   it's not a policy issue because lots of  countries have nuclear power they've had   very different policies the French have  been super gung-ho the Koreans um but in   no case the Koreans managed to get nuclear  power to come down at a rate of 1% per year   French have not managed to get it come down  and and in the US it's went up by a factor   of three so I could attribute the factor of three  to regulation but the fact that in almost exactly   the same period solar PV went down by a factor  of 10,000 and nuclear was at best flat tells   you there's very different Technologies people  are trying to understand what what is why is it   one things this way and one things the other way  well nuclear reactors are bespoke things they're   not really manufactured although the French and  the Koreans both tried to standardize them but uh   they're not very modular whereas things that come  down tend to be things like integrated circuits   that you just print or um that you can manufacture  in a highly modular way but the evidence for this   is still pretty loose and so it's hard to predict  AR priori from just characteristics actually the   main theory that gets quoted is one that J mcer  and I developed uh on and Sid uh uh where we you   know we built actually on earlier work by Stuart  Kaufman and Jose Lobo where they said what happens   if you just it actually goes back to John MTH who  was actually ironically the inventor of rational   expectations so also made a dart throw model of  technological change and uh so he was willing to   go either hand and um so Muth made the dart throw  model Jose and uh Stewart Etc improved it and we   improved it another notch we improved it so that  we could look at actually how is the modularity of   the technology affect the rates and we were able  to show that more modular technology should have   faster Improvement rates so a couple of people  have tested that the enough hand waving it's   true but it's not highly predictive because  the air bars are pretty big and modularity   is hard to measure and so forth yes question is  it possible to go back to where you compare the   different predictions you said one prediction  goes this way and the other one go yes so which   prediction did you want to look at graph oh is a  gra I don't remember which graph but I'll see but   but describe the graph a little more because  I I've got a lot of predictions on this gra   yeah it had a pink and a green pink black dots  pink and black dots okay let's see this one you   know yeah maybe this one yeah this these ones I  can't see anything what oh sorry that we're not   projecting anymore what happened what I think  Auto switches up well I can can remember this   is it easy could could somebody switch it back  on just what there a little the remote control okay it is a question about what was discussed  before yeah say your question out loud and maybe   eventually we'll get it we'll get we'll  get the camera back yeah so but what's   your question it's a it's a question of  what was discussed before which is um how to how to how to combine inertia first order  inertia with the S curve right which I if if   I underst right well the S curve already  incorporates the inertia but let me let   me ask a question maybe you were you meant to  ask which is there is some combination of these   two things and actually one thing we're  trying to do is to understand how how how   do policies affect things why is it not showing  the right screen it's St screen sharing it what   I think it's not screen sharing anymore yeah  okay it's also probably your second desktop so okay well we may just have to give up on this  because so unless somebody can come fix it because   I can't deal with it right now but but we are  trying to understand how effective our policies   and so we're collecting a lot of data to try and  look you know do the do the cause effect thing   what kind of policies produce what kind of effects  and so there's a serious problem to do but that   has to be combined with the inertial component  that can I ask my question sure yeah yeah it was   brilliant question that I would have asked well I  knew you would have so so um you made a point I I   will just describe what I saw okay yeah you made  a point that this curve here's the real curve and   here's the other curve and it's the difference in  years was eight years yeah and so if you think of   if you if you have these Technologies and it's  kind of first order inertia but then you're   trying to predict when they turn over saturate M  yeah you have to separate out I think the inertia   the first first order component of that and ask  when during developmental time do I gain enough   information to predict to responsibly or ACC yes  no that's yeah let let me I'm happy if you want   to no no go ahead finish the question right so  my qu so that's the L my point is that that the   question then becomes a long time when things  change at what time can you can you predict l   correctly right no that's one of the main things  we're doing in our paper and that and if you start   thinking about that you realize that L may change  you may have very little information about L and L   May so that becomes then a issue of of competition  of the co you know the costs actually so these are   it's kind of an assembly process that has to do  with the first order um diffusion things but also   with policy so I'm wondering if you even separate  out L about eight years and maybe just say say how   how can we kind of predict the future L whether  that's red yeah so I should have been a little   clearer we we are one of the key Innovations  we've made to solve this is to actually use   proper basian methods taking into account all  those effects I listed and working very hard   to get rid of this huge parameter bias but the  question we're asking is not just let's make a   prediction right now but as a function of where  we are how accurate is that prediction and those   aor bands that I showed you are saying right now  we think that's how accurate it is if we if I had   gone back and made a prediction 10 years ago the  air bars would have been much wider because we're   getting to the point where we can just begin to  detect the deviation from the exponential 10 years   ago it would have been invisible so so when the  paper comes out we'll be doing exactly what you   said now separating out the policy part from the  from the you know deterministic part It's Tricky   but but clearly when you look at the national  level you can see the bumps and the curves caused   by the policies and we want to understand better  how to influence those bumps in the curve so the   opposite reading of your talk is that all these  terrible predictions we show us didn't matter   because investors and scien got together and acted  in a way that in effect falsified them and yet you   seem to be expending a large amount of energy on  making better predictions so why why because go   back to what I said about Moore's Law and Ali  Ray Smith it was critical that he could say in   five years investors you I'm asking you for 100  million bucks finish this thing we've got all the   technology in place mors law will deliver it for  us in five years that's huge okay and and in fact   I would the you know the fact that the predictions  have consistently underpredicted deployment of   Renewables you know predicted excessively high  costs for the last 20 years has slowed them down   if it hadn't been for that we'd have more than we  have now so so I'm not saying everything is you   know just pre-ordained and it's going to happen  regardless of what we do at all it really does   matter but but the technology and as you know  and and and its usefulness and so on dictate a   lot about how things are going to go then nice to  see it's I mean I would I would say that probably   a lot of it depends on actually articulating a  mechanism because otherwise people just have to   have their faith yeah yeah so that when you can  refer back to some whatever segmented production   cycle uptake all these things that that's  the thing that convinces people yeah just   don't smart but I don't know yeah so I mean you  know that but I no I understand think back when   we were seeing the logistic winning we thought  great I can explain that you know I had a nice   little story there's I wrote a draft there's two  paragraphs with a mechanism but if Gumpert it's   as really better as it suggests it is although  right now what we're just doing right now is   doing the head-to-head forecasting competition to  see which one forecasts more accurately GTS is a   little bit tricky because it's a stiffer equation  in some sense and so um but if it's gumper it's   then I'm a little puzzled know if if if it was  logistic I could tell a story that's kind of like   bacteria you know for instance on the consumer  side the more people have it the more people see   the oh your friend has it now I need to get one um  but you saturate the market at some rate one minus   X so it's kind of the same story as the bacteria  yeah I mean I guess one thing I wonder about is   how much of the underlying Innovation processes  production processes kind of washes out in   the noise like like what's the princip I mean  yeah you know much better than I like generic   principles of Technology development that have  some hope of adoption a gain a foothold had you   know again depending on what kind of Technology  it is how does that feed back um into the cycle   uh you know in in this sense transistors are  kind of easy yeah um yeah whatever the physics   of it was pretty well understood um but here  there I I think I mean there are a lot more   moving parts and to the extent that you could  abstract different Industries but du Tails back   yeah be interesting I don't well because it's it  depends on so many other things so it's as you   say the grid is completely essential what and  as talks about it's complex but you've made a   decision that in a way from what I hear because  you're saying well the the the models are wrong   and hence people are not investing so you want  to have the right models so that people invest   in in in which is great but it's still an opp  cost it still means they might want invest in   other things yeah so you're making a decision that  that's what's worth investing on but it means that   there are other things that could be invested  by all means happen so is it NE by all means   and let me just tell you the thi

2024-08-19

Show video