Machine learning in the built environment

Occasionally I like to write about other things besides just walking and eating in the mountains. This was originally written for a course in my master’s of business analytics program. Saving here for posterity.


New Data, Old Habits: The Promises and Pitfalls for Driving Sustainability Outcomes with Data & Machine Learning

In 2018, Jim Gao saw his baby come to life. Not literally, mind you, but it’s not a stretch to imagine the amazement was similar. One day, Jim and other Google engineers cut energy used for cooling at one of their hyperscale (i.e., very large) server warehouses by 40%, and 15% of overall energy overhead, without changing a single physical piece of machinery (Figure 1). Instead, they turned over complete control of the facility’s energy use to a machine learning algorithm. With a single click (and many preceding months of hard work), they reduced energy use and accompanying greenhouse gas emissions from a facility that already led the industry in energy efficiency. Such is the power of leveraging data & machine learning (ML) for combating climate change, our greatest challenge of all.

The purpose of this blog is to use two applications of data & ML to explore both their promise and potential shortfalls in catalyzing sustainability efforts. First, I’ll examine how data & ML can revolutionize energy efficiency in the built environment (human-made structures such as buildings, homes, and offices). Secondly, advances in both physical and digital technology for tracking deforestation hold promise for ameliorating environmental degradation in the palm oil industry’s notoriously abusive supply chain. Yet, this area provides a vivid example of why technological advancements alone are insufficient to drive this positive change; changes in the social forces that surround technology are equally, if not more, important. Together, these examples show how data & ML offer much promise for driving impressive sustainability outcomes, yet their impact can be pushed in unintended directions, reduced, or eliminated outright by powerful social forces such as economic incentives, political opportunism, and corruption.

Figure 1: Energy use at a Google data center with and without the AI energy management system enabled. PUE stands for Power Usage Effectiveness, an industry-standard efficiency metric calculated by dividing the total energy used in a facility by the energy consumed specifically for information technology activities. (source: Google)

AI in the Built Environment

Jim Gao latched on to the idea of creating a ML energy management system for Google’s data centers because reducing energy use at these large facilities is no trivial concern. These massive collections of servers power vast amounts of the Internet that we rely on, such as Google searches, Youtube videos, Gmail, and cloud computing applications. According to the Department of Energy, these centralized data centers are estimated to account for 2% of global energy use (Google, n.d.), reflecting society’s increasing reliance on the Internet and cloud computing. Yet, while global demands on cloud computing at these large data centers have increased by 550% from 2010 to 2018, aggregate energy use has increased by only 6% over that time frame (United States Data Center Energy Usage Report, 2016). While improvements in physical technology such as more efficient fans and cooling machines contribute to this, advancements in using data and machine learning play an outsized role.

A data warehouse’s energy system consists of a multilayered, intricate dance of vents, fans, ductwork, power lines, and cooling towers all working together to either provide power to or cool down enormous racks of servers. It is an illustrative example of a complex system in the vein of how systems thinkers use the term: a system characterized by nonlinear feedback loops, uncertainties, and unintuitive emergent reactions borne from billions of possible system configurations between internal components and external factors such as weather and the energy sources the system pulls electricity from. Such complex systems are ideal cases to augment human intuition with machine learning that can crunch the numbers of different scenarios and find the ideal system configuration for any given moment — a computational challenge that human brains simply aren’t wired for.

Human judgment doesn’t go completely out of the picture, however. Google engineers continuously monitor the AI system’s present actions and future recommendations and have veto power over any move they deem too risky or disruptive. In fact, the first recommendation from a prototype iteration of the system was to shut down the entire warehouse itself (Google, 2019) — after all, that would be incredibly energy efficient! Human intuition is also highly valuable in optimizing the system during extreme situations, say, a tornado, that are outside the AI system’s training.

In addition to optimizing system configuration, AI is revolutionizing energy use in large data centers by also altering the timing of energy use. In many places, electricity that flows into the energy grid is produced from a variety of sources — from carbon-intensive coal plants to renewable sources like wind and solar farms. And since there are times when the wind is not blowing or the sun is not shining, the carbon cost of a unit of energy drawn from the grid varies throughout the day. Google, for instance, has developed machine learning algorithms that schedule non-urgent computing tasks (like software updates or Youtube video processing) to times when low-carbon power sources make up higher proportions of the energy grid (Radovanovic, 2020), thus decreasing carbon emissions without sacrificing productivity or service quality.

There is also strong potential for AI-driven energy management elsewhere in the built environment, and it is sorely needed. Despite pledges to cut emissions, economic incentives, and public pressure, residential and commercial buildings still account for roughly 39% of U.S. carbon emissions — with no downward trend in sight (Lai, 2020). And while commercial buildings are broadly more efficient than residential buildings (Figure 2), the sheer scale of the commercial built environment means there is still plenty of opportunity to mitigate operational energy use.

Figure 2: Direct and indirect GHG emissions from the commercial and residential building sectors. “2014: Buildings. In: Climate Change 2014: Mitigation of Climate Change. Contribution of Working Group III to the Fifth Assessment Report of the Intergovernmental Panel on Climate Change” (Lucon et al., 2014)

After moving its data centers to fully-autonomous AI energy management in 2018, Google is now looking to offer these capabilities as a service to other commercial facilities that could also reap the efficiency benefits. Potential clients for this service would seem to fall into two categories — those whose sheer scale makes AI-driven energy efficiency cost-effective, and those with sufficiently complex energy management or production systems. In the scale category, AI drives marginal efficiency gains relative to human management through its ability to continuously make small adjustments in response to conditions; these adjustments are made many thousands times more often than what human operators would make given identical conditions. And in the complexity category, AI systems can deliver superior performance even in small-scale applications through their ability to model and predict many billions of system configurations and adapt to emergent and unintuitive outcomes faster and more comprehensively than human intuition alone. Of course, the greatest efficiency gains likely lie with clients who combine both scale and system complexity — like auto or aerospace manufacturers, large government buildings like the Pentagon, and distribution centers.

Another example of AI in the built environment comes from McKinstry, a Pacific Northwest-based sustainability innovator that drives energy and materials efficiency for new and retrofit buildings. According to Devon Powell, a Program Manager for McKinstry, the company uses an AI-driven energy management system with some of their clients’ more complex commercial buildings that optimizes energy flow between different components based on occupants’ energy use and outside weather patterns (Powell, 2021). McKinstry also provides systemwide auditing and project prioritization services for clients with a portfolio of buildings and potential sustainability projects, like commercial developers and school districts. Mr. Powell states that many of the school districts his firm works with have historically prioritized routine maintenance and energy efficiency projects alike with rather inefficient, anecdotal approaches where the squeaky wheel gets the grease. Conversely, McKinstry uses data on individual equipment, energy use, and buildings across an entire district to prioritize projects with highest ROI and environmental impact.

Both Google and McKinstry demonstrate the positive impacts of fostering data-driven cultures. The trajectory of data-driven innovation is clear to see in how Google has rolled out their AI energy management, beginning in 2014 with an AI recommendation system at one facility where engineers then implemented all recommendations manually, moving to an autonomous AI management system in 2018 across all data centers, and then building on those successes in 2021 by looking to productize the innovations as a SaaS platform. At every step of the way, Google invested in training and adoption for its on-site data center engineers, ensuring they were empowered by the system rather than threatened. McKinstry similarly finds success in fostering data-driven decision making and cultural change with its clients by building trust over time; snowballing the success of smaller projects (like adding renewable energy or upgrading equipment in a single building) into larger, more impactful ones (like a portfolio-wide audit of a school district’s assets & co-developing the district’s sustainability strategy).

Big Data in Supply Chain Management

On the other side of the world from Google’s and McKinstry’s energy efficiency efforts in the United States, a European Space Agency Sentinel 1-A satellite floats in low Earth orbit above the Indonesian jungle. Unbothered by rain, darkness, smoke, or cloud cover, onboard radar systems map the landscape below in extremely precise 30 meter by 30 meter blocks. The data is continuously fed back to a Google Cloud-based system and publicly available to any partner organization in the RADD (Radar Alerts for Detecting Deforestation) system. Prior to this system, state-of-the-art satellite imagery used optical sensors that couldn’t penetrate cloud cover; for tropical environments where clouds or haze predominate this could mean several months would go by before usable images could be obtained. The resolution was lower, as well, with radar-based resolution roughly nine times as precise as optical sensors. The result is now a system that obtains accurate imagery with every orbit no matter the atmospheric conditions, and generates publicly-available alerts that can help spur greater transparency and sustainability in supply chains.

The common phrase “you can’t manage what you don’t measure” is more appropriate here than anywhere. For the consortium of ten of the largest global palm oil companies involved in the program, this kind of accurate, highly responsive supply chain information marks a major breakthrough in managing for social and environmental sustainability. Yet, utilizing more sophisticated technology is only part of the story. As with so many other sustainability challenges, the people side of the equation far outweighs the technological challenge. More data does not always equal more change. In recent decades, palm oil has grown into a massive commodity in global trade through its versatility and cheap production costs. It’s found in roughly half of all household and food products sold in developed countries (Environment News Service, 2020), including products such as shampoo, toothpaste, laundry detergent, frozen foods, cookies, peanut butter, margarine, skin lotion, and cosmetics. But its production comes at a cost. About a quarter of global carbon dioxide emissions can be attributed to the agricultural sector, with most of those emissions due to clearing land for crops. From 2002 through 2019, global annual tropical forest loss averaged 8.3 million acres — an area larger than Maryland (Global Forest Watch: Dashboard , n.d.). And while studies have found differing impacts of palm oil production on overall deforestation, a 2020 study (Meijaard, 2020)[1] finds that across 23 studies spanning from 1972 to 2015, palm oil is a significant driver of deforestation (Figure 3).

Figure 3: Oil palm’s estimated role in deforestation aggregated across studies, years, and regions. Panel A depicts the contribution of oil palm to overall deforestation, while B shows the percentage of all oil palm expansion that cleared forest (Supplementary Methods). There were no data for Peru and South and Central America for panel A, and no global data for panel B. Southeast Asia (SE Asia) excludes Indonesia and Malaysia, which are shown separately, while South America excludes Peru. Each filled circle represents one time period from a single study, with individual studies represented by distinct colours. The size of the circle corresponds to the relative number of area years represented in that time period (larger circles represent a larger study area and longer time period of sampling). Boxplot middle bars correspond to the unweighted median across study-time periods; lower and upper hinges represent the 25th and 75th percentiles of study-time periods; and whiskers extend from the upper (lower) hinge to the largest (smallest) value no further than 1.5 times the interquartile range from the hinge.

 Despite increasing awareness among buyers about the deforestation, habitat and species loss, air pollution, and human exploitation that can accompany palm oil production, there is little appetite to pay a premium for certified sustainable palm oil. Simon Lord, chief sustainability officer at Sime Darby, largest palm oil producer by acreage in the world, attests that “buyers don’t want to pay for [certified palm oil]” (The Asean Post, 2019). He states that, additionally “there is increasing resentment among growers that the other actors in the supply chain are not stepping up.” This low demand for sustainably harvested palm oil means Sime Darby can only sell about half of its certified palm oil at a premium price; the rest gets mixed in with conventional oil and sold at standard prices.

Additionally, palm oil production exists within a complex web of politics and capitalism where regulations are nonexistent or infrequently enforced, land rights of marginalized groups are abused, and workers exploited (The Asean Post, 2017). Mikaela Weisse, who leads GFW’s strategy and partnerships for satellite-based forest monitoring, states that “the biggest things we’ve found, while there are ways we can improve the data, the real blocks are institutional challenges from hard-to-get-to forest areas, governance, and corruption, [all of which are] really big problems” (Coca, 2019). There is credible evidence that big, splashy announcements like the RADD system are little more than greenwashing. According to a report from the World Wildlife Fund, not a single company that made pledges for 100% sustainable palm oil production in 2010 during the inaugural Consumer Goods Forum (Preston, 2010) has met its commitments.

Additionally, the report highlights that the challenge and the failings are on the social side, not the technological; of the 173 companies surveyed, only 27% require their suppliers to have deforestation-free policies, only 10% require traceability to mills and plantations (World Wildlife Fund, 2020). Even more telling, a staggering 24% of companies did not respond at all, perhaps signaling resistance to even basic transparency regarding their business practices. And underlying all this, global forest loss is still increasing; despite the Covid-19 pandemic, tropical forest loss increased 12% in 2020 compared to 2019 according to a 2021 Global Forest Watch report (Mooney, Dennis, & Muyskens, 2021).

These social challenges point to a truth about data and technology. They can open new frontiers in how we make sense of the world, but they are only as good as how we use them. Getting the political, social, and economic forces that surround data and technology right is far more important than getting the technical specifications correct. As Annisa Rahmawati, senior forest campaigner at Greenpeace Indonesia, so succinctly states, “No additional technology is needed [to socially and environmentally clean up supply chains]. Technology is one thing, implementation is another” (The Asean Post, 2019). If we do not pay attention to these forces in how we use technology or interpret data, we are only perpetuating the injustices and abuses of the past under the guise of new platforms and methods.

Tying It All Together            

In many ways, we capture more data about the world around us and have more computing power to analyze that data than ever before. In 2018, Forbes estimated that more than 90% of the world’s online data had been created in the previous two years (Marr, 2018), and this trend has only accelerated since then with subsequent advances in technology and adoption in nearly every facet of life, from consumer smartphones to smart sensors attached to vents in a hyperscale data warehouse. This makes the question of how we harness that data more critical than ever before. If we want any chance of reducing greenhouse gas emissions enough to get on a global trajectory that avoids the worst climate change we’d do well to remember that our greatest challenges rarely ever lie in external breakthroughs, they lie within and between us.



Sources

Coca, N. (2019). Data ≠ Change. Thanks to satellites, we can monitor our forests better than ever before. So why is global deforestation still increasing? Retrieved from Earth Island Journal: https://www.earthisland.org/journal/index.php/magazine/entry/data-

Environment News Service. (2020, April 27). Palm Oil Industry Funds Radar to Detect Deforestation. Retrieved from https://ens-newswire.com/2020/04/27/palm-oil-industry-funds-radar-to-detect-deforestation/

Global Forest Watch: Dashboard . (n.d.). Retrieved from https://www.globalforestwatch.org/dashboards/global/

Google. (2019, September). Machine learning finds new ways for our data centers to save energy. Retrieved from Google Sustainability: https://sustainability.google/progress/projects/machine-learning/

Google. (2019, September). Machine learning finds new ways for our data centers to save energy. Retrieved from Google Sustainability: https://sustainability.google/progress/projects/machine-learning/

Google. (n.d.). Data Center Efficiency. Retrieved from https://www.google.com/about/datacenters/efficiency/

Lai, J. &. (2020). Review on carbon emissions of commercial buildings. Renewable and Sustainable Energy Reviews.

Marr, B. (2018, May 21). How Much Data Do We Create Every Day? The Mind-Blowing Stats Everyone Should Read. Retrieved from Forbes: https://www.forbes.com/sites/bernardmarr/2018/05/21/how-much-data-do-we-create-every-day-the-mind-blowing-stats-everyone-should-read/

Meijaard. (2020). The Environmental Impacts of Palm Oil in Context. In preprint.

Mooney, C., Dennis, B., & Muyskens, J. (2021, March 31). Global forest losses accelerated despite the pandemic, threatening world’s climate goals. Retrieved from The Washington Post: https://www.washingtonpost.com/climate-environment/2021/03/31/climate-change-deforestation/?itid=

Powell, D. (2021, March 26). Senior Project Manager, McKinstry. (N. Bender, Interviewer)

Preston, S. (2010, December 1). Consumer Goods Forum plans to tackle deforestation and other key drivers of climate change. Retrieved from The Guardian: https://www.theguardian.com/sustainable-business/consumer-goods-forum-deforestation-climate

Radovanovic, A. (2020, April 22). Our data centers now work harder when the sun shines and wind blows. Retrieved from Google Blog: https://blog.google/inside-google/infrastructure/data-centers-work-harder-sun-shines-wind-blows

The Asean Post. (2017, December 24). The palm oil fiefdom, Part 1: Indonesia reborn. Retrieved from https://theaseanpost.com/article/palm-oil-fiefdom-part-1-indonesia-reborn

The Asean Post. (2019, November 15). Can Radar Tech Clean Up The Palm Oil Industry? Retrieved from https://theaseanpost.com/article/can-radar-tech-clean-palm-oil-industry

The Asean Post. (2019, January 14). Sustainable Palm Oil: No One Wants It. Retrieved from The Asean Post: https://theaseanpost.com/article/sustainable-palm-oil-no-one-wants-it

United States Data Center Energy Usage Report. (2016). Retrieved from https://eta.lbl.gov/publications/united-states-data-center-energy

World Wildlife Fund. (2020, January). Palm Oil Buyers Scorecard. Retrieved from https://palmoilscorecard.panda.org/file/WWF_Palm_Oil_Scorecard_2020.pdf

Leave a Reply

Fill in your details below or click an icon to log in:

WordPress.com Logo

You are commenting using your WordPress.com account. Log Out /  Change )

Twitter picture

You are commenting using your Twitter account. Log Out /  Change )

Facebook photo

You are commenting using your Facebook account. Log Out /  Change )

Connecting to %s