Text Mining Public Perception of Smart Meters in Victoria

Today the Herald Sun ran a story proclaiming that smart meters are here to stay and invited their readers to comment on whether the government should scrap the smart meter program. I am not going to comment here on the journalistic quality of the article but concentrate on comments section which gives stakeholders some valuable insight into the zeitgeist of smart metering in the Garden State.

By applying an unstructured text mining application I have extracted the key themes from the comments on this story. When analysed in conjunction with the structure and content of the story, we get some interesting insights into public perception.

To start with I excluded the words “smart”, “meter” and “meters” in order not to be distracted by the subject under discussion. This is what I got.

Word clouds often seem to point to a collective meaning that is independent of individual attitudes. If this is the case then the strong message here which we could interpret as a collective rejection of what is seen as government control being favoured over the wishes of the “people”. This may be more of a reflection of the Herald Sun readership rather than a general community concern however.

If I remove “government” and “power” we get a closer look at the next level of the word cloud.

An aside of note is that we see that Herald Sun readers like to refer to the premier by his first name which is perhaps a sign that he still has popularity with this demographic.

One interesting observation to me is that despite its prominent mention article, the myth of radio frequency radiation from smart meters is not a major concern to the community, so we are unlikely to see a repeat of the tin foil hats fiasco in California.

Once we get into some of the word cloud detail, we see the common themes relating to “cost of living”, namely the additional costs to the electricity bill of the roll out and potential costs associated with time of use pricing. The article does mention that time of use pricing is an opportunities for households to save money. Time of use pricing is also a fairer pricing regime than flat tariffs.

The other important theme that I see is that the smart meter rollout is linked to the other controversial big technology projects of the previous Victorian government – Myki and the Wonthaggi Desalination Plant. But the good news is that the new government still has some cache with the public (even in criticism readers often refer to the premier by his first name). The objective now should be to leverage this and start building programs which smart meter initiatives which demonstrate the value of the technology directly to consumers. This in part requires unlocking the value of the data for consumers. I’ll speak more about this in future posts.

UPDATE: For interpretation of word clouds I suggest reading up on concept of collective consciousness.


Appliance Penetration and the Wisdom of Crowds

Some of the burning questions for electricity utilities in Australia have to do with appliance take up. I decided to see what the wisdom of crowds could tell us about the take-up of some key appliances which are affecting load profiles and consumption trends. My crowd-sourced data comes from Google Insights for Search. I have taken the weekly search volume indexes for three search terms: “air conditioner”, “pool pump” and “solar pv”. In addition, I also took the search volumes for “energy efficient” to see if there has been a fundamental change in the zeitgeist in terms of energy efficiency.

Firstly, let’s have a look at Google “air conditioner” search data.

The graph shows strong seasonality with people searching more for air conditioners in summer which makes sense. We see indications of how profound the growth of air conditioners has been in Australia (and South East Queensland in particular), I decided to compare growth in air conditioner searching by country and city. Since 2004, Australia ranks second behind the US for air conditioner searches. For cities, Brisbane and Sydney rank fourth and fifth in the world, but if we adjust for population they rank second and third respectively behind Huston. This has been one of the causes behind the recent difficulties in forecasting demand. One of the big questions is will air conditioning load continue to grow or has air conditioner penetration reached saturation point? Read on to find some insights that I think this data may have uncovered.

When we look at the data for the search term “energy efficient”, we get the opposite temperature effect with dips in searches during summer and maybe winter is noticeable in this graph.

This tells us that people become less concerned with energy efficiency as comfort becomes more important which has also been shown in other studies. But if we want to look for underlying changes in behavior then we need to account for temperature sensitivity in this data and the first thing we need to do is come up with a national temperature measure that we can compare with the Google data. To do this I get temperature data for Australia’s five largest cities from the Bureau of Meteorology and create national daily maximum temperatures for 2004-2011 comprised of a population weighted mean of the maximum temperatures of the five largest Australian cities. This accounts for about 70% of Australia’s population and an even greater proportion of regular internet users. Now we can quantify the relationship between our appliances, energy efficiency and temperature.

Below are the scatter charts showing the R2 correlations. “Solar PV” is uncorrelated with temperature but all of the other search terms show quite good correlation. You may also notice that I have tried to account for the U-curve in the relationship between “Energy Efficient” and temperature by correlating with the absolute number of degrees from 21C. The main relationship is with hot weather; accounting for the U-curve only adds slightly to the R2. Interestingly, people don’t start searching for air conditioners until the temperature hits 25C, and then there is a slightly exponential shape to the increase in searches. For the purposes of this post I will stick to simple linear methods, but further analysis may consider a log link GLM or Multiple Adaptive Regression Splines (MARS) to help explain this shape in the data.

Now to the central question that this post is trying to answer: what are the underlying trends in these appliances, can we find this out from Google and BOM data and can we get some insight into current underlying trends and how this might help uncover the underlying trends in consumption and load factor.  To do this I create a dummy variable to represent time and regress this with temperature to see to what extent each factor separately describes the number of Google searches. I build separate models for each year which separates the trend over time in searches from the temperature related ones.

But before I do that I can look at the direct relationship between annual “solar PV” trends.

There was not enough search data to go all the way back to 2004 (which is of itself interesting) so we only go back 2007. What we see is a large growth in searches in 2008, statistically insignificant trend in 2009 and 2010 and a distinct decline during 2011. It looks like removal of incentives and changes to feed in tariffs are having an effect. The error bars show the 95% confidence interval.

Now on to pool pumps. Here we see a steady rise on searching for pool pumps which indicates that we can expect growth in pool pump load to also grow nationally. If anything it looks like the search rate is increasing and maybe apart from 2008 maybe not affected by the 2008 global downturn.

Once we account for temperature variability we see really no trend in terms of energy efficiency until 2010. This came after the collapse of Australia’s carbon trading legislation and the collapse of internally accord on climate change policy. It stands to seems to me that this is also reflected in the public concern with energy efficiency. It seems to me that if there is widespread public concern about the contribution of electricity to cost of living then it should be reflected here but it isn’t. This also seems to suggest that for consumers the motivation towards energy efficiency is driven by a sense of social responsibility rather than being an economic decision.

Finally, air conditioning. What we see represented here is the rapid growth in air conditioning that happened in 2004-2005 with a slowing in growth from 2006-2008. It looks like maybe the government rebates of 2009 may have been partially spent on air conditioning. But what we see is that from 2010 onwards there has been no significant trend in search term growth. Does this suggest that we have finally reached saturation some time during 2010?

What the water sector can learn from analytics

Deloitte and the Australian Water Association has released the 2011 State of the Australian Water Sector Report and the most important issue raised by the industry is that of “sustainability” and some differences of opinion on whether environmental or economic sustainability is the more important concern. An analytical approach can help inform the industry on these important sustainability issues in different ways. The “smart grid” is an important factor in this to the extent that it is just the technological enhancements that will come with natural replacement of current infrastructure and therefore nothing particularly mystical. As I have said before, the “smart grid” is really just the slow evolution of old measurement technology and the “smart” part is how we get better at extracting useful insight from the data.

A case in point is climatic variability. I am studiously avoiding use of the term “climate change” because the political debate around this is really not useful in the context of managing climate-sensitive resources. The fact that our climate is highly variable is beyond debate and I will leave it for other forums the debate whether that variation is directional, cyclical or some combination of the two.

One thing that the water industry can learn from the electricity industry is around the work done to understand weather related demand and how to account for that variability. I have said before that while I do not believe the electricity utilities have yet cracked how to properly account for weather-related variability of demand, they have done a lot of work in this respect that could yield insights for water utilities. If this variability can be properly understood then we can isolate underlying growth factors and develop consumption scenarios under different hot dry climatic scenarios. From an analytics point of view, if we can model these down to individual consumers then we can develop incredibly rich scenarios with different cohorts of population responding in different ways. To this extent all utilities would do well to turn to scenario modelling rather than traditional forecasting in order to better understand the underlying growth in demand and provide a solid methodological basis for informing the policy debate (I’ll talk more about the difference between forecasting and scenario modelling in a future blog post).

In terms of economic sustainability, pricing and price setting is the key analytical exercise. Understanding price and demand elasticity is the critical element in developing future economic sustainability of the water industry. This is still a way off for water but is worth considering because it can help target spending on network infrastructure renewal so that the right data is collected for future modelling. Usually, elasticity is expressed as an average for all users. What is far more important to understand is the distribution of elasticity with a given population and whether there are other factors that describe elasticity segments. This can help drive product differentiation and demand management strategies which in turn support the economic sustainability of the network.

Is there a moral argument for time of use pricing?

I recently came across an interesting paper with a novel argument for demand pricing.  In a previous post I explained why peak demand drives network costs. Because we mostly have flat tariffs in Australia we have the situation of cross subsidy whereby people with ‘flat’ demand profiles subsidise those with ‘peaky’ demand profiles. Consider this example.

From a policy point of view there is nothing wrong with cross subsidy per se, but it is important to know who are the winners and losers in the transfer of costs. If the flatter demand belongs to lower income consumers and peaky demand belongs to higher income consumers then flat tariffs subsidise the rich and transfers demand costs to the poor. If this is the case then it is hard to argue that a flat tariff structure is fair.

To test this there are a number of factors that I have considered:

  1. How should ‘peakiness’ be measured?
  2. Is there an inherent link between ‘peakiness’ and household income?
  3. Therefore is a flat tariff fair or unfair, based on who is being cross-subsidised?

What I would like to do here is present a detailed analysis but I cannot do this because most of the data I have is highly confidential and there is a paucity of public data available. What I can do is share some general observations based on my experience across a number of jurisdictions and my general approach.

The standard measure of peakiness is load factor: for a given period this is the maximum demand measure divided by the average. This gives a measure of peak relative to the underlying demand. In the example above the ‘peaky’ profile is about 3% peakier than the flat profile. But another measure is the actual range in demand between the peaky profile and the flat profile. In the same example we get a 60% difference between the peaky and flat profiles.

What I have noticed is that if we look at load factor this is more or less uncorrelated with household income if I remove controlled load (the rationale is that controlled load is off peak anyway and controlled by the distributor, not the consumer).  If I look at the range in demand between the trough and the peak, we get a reasonably strong correlation with income. But we also have a correlation between income and total consumption, so range could just be a function of total consumption? That is, if you use more then the variation between your peaks and troughs are also going to be larger.

So what are the conclusions?

Firstly, I can’t find any evidence that demand pricing will inherently transfer costs from wealthy households to poor households. Wealthy households by virtue of their greater consumption contribute more to peak demand even though they are not inherently peakier in their usage profile.  From my analysis I can’t say that poor households are inherently flatter in demand and therefore subsidise richer households, but what I can say is that given an equivalent household income, flatter demand households do indeed subsidise peak use households under a flat tariff structure and that demand pricing such as a properly designed time of use tariff could remove this cross subsidy.