Lifetime Learning for AI Everywhere

For decades, from even before we called everything the IoT (Internet of Things), maintenance has been the barrier to digital sensing and operating of the physical world. Wired sensors were reliable, but expensive to install, and often an esthetic nightmare once installed. With self-power, sensors became cheap enough to put everywhere, but faced a new challenge—maintenance.

For a long time, deployments were limited by battery life. Many initiatives were short lived, running until the batteries wore out. Changing the batteries was expensive, sometime more than the initial installation. Committed organizations developed scheduled battery changes to control costs, just as they had done before for re-lamping projects.

We solved that problem by making sensors so cheap we could just leave them and install replacements. Or we (notably members of the EnOcean Alliance) tuned communications to be so light-weight that in situ energy harvesting could keep systems working.

Now we face yet another maintenance challenge, that of intelligence management.

Today’s sensors have become smarter, sometimes referred to by the indeterminate name “edge devices”. Sensors and Edge Devices likely transmitted more than 20 zettabytes of data for central storage last year, although there are no firm estimates on 2017 data gathering. With that much data being stored, the communications requirement was easily in yottabytes.

This much data creates a new challenge. The IoT not only requires that we get actionable information that matters, but that we get it before it is too late to matter. There is too much data and too many situations to rely on timely central decisions.

The enable drinking this firehose of data, we are starting to rely on sips at the edge. Edge Devices are making the initial decisions as to what data means, and what data needs to be brought into the middle. Local decisions are made faster, without interference from temporary high priorities elsewhere in the IoT. For all but the simplest scenarios, this model requires learning at the edges. There are large open source libraries now of Artificial Intelligence (AI) code for Raspberry Pi and Arduino.

This presents a new maintenance problem, managing and updating AI routines and algorithms.

The big software companies are preparing the tools we will need. Thousands of AI systems in each building will require tools to manage the rapid evolution algorithms. New algorithms will require managed roll-outs Rapid evolution forces diversity of algorithm and information as systems will change far faster than their installed life. Oracle is pushing GraphPipe, an open source software project for efficiently deploying and managing AI models at scale. Microsoft is right there with them, with large platform management announcements expected this Fall.

The problem of managing intelligence in millions of devices is solved already, before most people know they have the problem.

In the last year Pi architecture devices have blown right past the $40 and even $20 price points, with full systems expected for $7 and perhaps $4. Arduino platforms not only run open source Linux and Android, but with open source hardware offer potential easy integration directly onto integrated specialty hardware components.

The barriers to fully intelligent small systems across every aspect of buildings are falling even faster than pioneers such as Alper Üzmezler and the Project Sandstar for smart controls have publicly projected. It is my personal belief that while full platforms such as the Pi have higher initial costs, in part because they include a GPU not needed to manage a display, that this means they are pre-adapted for high speed signal processing. This will not play out slowly.

Transactive Energy and Farm to Plug

I just got back from the Third International Conference and Workshop on Transactive Energy in Portland. There is wide consensus on the inevitability of transactive energy even as there are struggles as to how to get there.

Transactive energy was initially conceived of as a way to set spot market prices for electric energy (power) during times of peak demand or temporary supply shortfall. Transactive energy is based on the path-breaking research of Clearwater and Huberman at the Xerox Palo Alto Research Center (PARC) published in 1993. At PARC, they created moment-by-moment thermal markets to manage data center cooling; an agent on each server bid for the cooling it needed. This approach eliminated hot spots and reduced energy costs even as it eliminated the need to develop ever more complex control and sensing strategies.

Distributed energy makes the problems of effective grid operation worse. Distributed energy refers to the developing model in which every node on the grid is potentially a power source as well as a power user, driven largely by renewable energy such as solar photovoltaics (PV) and wind. Distributed energy changes the centrally managed, essentially hub-and-spoke distribution model in which energy flows down into what is potentially a two-way peer-to-peer network over the same infrastructure. Sites which contain Distributed Energy Resources (DER) can choose whether or not to come to market at any moment. Transactive energy is the developing means to manage this growing complexity.

Distributed energy is local, so distributed energy markets (and prices) must be local. Traditional local prices in power, referred to as locational marginal pricing (LMP) or nodal pricing is based on physical limits of the transmission system—a single bottleneck can affect all “downstream” points. LMP can be set centrally, calculated based on line physics and historical use. DER potentially places the power sources downstream of the congestion, and alongside the power customers. Nodes containing DER can decide whether the energy available is used to support the grid or internal purposes. Only actual markets and set clearing prices for DER.

There is no effective ownership of DER without local storage. Without local storage, grid nodes are always price-takers. Grid operators have a strong and legitimate interest in throttling how much DER is dumped onto the grid at any moment. Without local storage, grid operators must be able to turn off DER, i.e., set when a node can come to market. Even if a node invests capital in DER asset, if a third party determines what prices the node must take for the product of that asset, and controls when that asset can come to market, then the owners of that node cannot be said to own the asset.

Local markets will not really work without local storage. Local storage is necessary to create actual economic ownership of DER.

The best use for DER is and will always be local consumption. A building need not be Net Zero Energy (NZE) to consume power locally first. Use energy locally first. The next best use for DER is to store energy locally, perhaps for later consumption on site. Any excess, or any deficits in local power can then be made up through market operations. This is the essence of the new power movement, sometimes called Farm-to-Plug.

Start with a Zombie Fortress

In smart energy, it is easy to get distracted by utility incentives and demand response and other tariffed actions. Utility tariffs are set in stone months or years before an actual set of market conditions arise. Demand Response events miss the supplier’s pain-points while ignoring opportunity for the building owner. “Running a meter backward” is a silly demonstration project that works only so long as very few people do it. All of these are regulatory fantasies that violate the laws of economics and physics. For a smart energy engineer, it is better to start with a more realistic fantasy.

Smart Energy starts with a Zombie Fortress.

Many today who are uneasy about politics and culture and technology dream of a place to get away if things fall apart.  Zombies have no politics, no ideologies. They are mindless, and ugly, and the perfect nightmare for a time when any judgment potentially offends. The coming Zombie Apocalypse is the perfect non-specific eschatology for our time.

The Zombie Fortress is where you go to be safe from the world. Folks can share their desire for a Zombie Fortress without getting into discussion of politics with their friends. The Zombie Fortress names a non-political escape, a bolt-hole to go when everything goes wrong. (Some might claim that the editor of Automated Buildings has retreated to a Zombie Fortress.) Plans for a Zombie Fortress cannot assume that the grid will work, or that the neighbors will be a useful source of supply or resilience.
The challenge of the Zombie Fortress is to live a full life within the site-generated power. System efficiency is critical, certainly, but it is swamped by the power usage efficiency; the operating margin must go as close to zero as doable. This means no power spikes, and no wasted power. Systems must be negotiate so that intermittent systems do not run at the same time. Any extra power, moment to moment, must be pre-consumed or stored.

Above this is a policy layer. If you habitually use power into the night, that is the basis for the power storage goals. Weather reports may set to pre-consumption goals. Systems must decide how important they are and run, or not run, accordingly. Engineers will be in short supply after the Zombie Apocalypse, so the systems in the fortress must integrate themselves.

But maybe the burning times have not yet come. For now, you decide to use the Zombie Fortress as your Party Pad in the in the mountains. Maybe the Fortress cannot produce enough power each day to keep the lights on, the water pumped, and the environment comfortable during sustained use. If the Fortress plans, if it it stores power all week, though, it can support a two day weekend. Maybe a three-day weekend requires two weeks of storage.

But you want to throw a big party. The last party was automatically base-lined by the Fortress. You contact the Fortress from afar, and ask when it will be ready. The Party Pad / Fortress informs you that it will need four weeks to accumulate enough stored energy, five if you send in a cleaning crew during the week in advance. This is the right level of owner interaction.

Transactive energy within the fortress is the simplest integration strategy devised. Traditional integration requires detailed knowledge of all systems, solving what economists call the knowledge problem. Transactors don’t need knowledge of their trading partners, merely common agreements. New systems must merely introduce themselves to the market. Each system, to participate competently in the market, needs to understand its own patterns of use and load shapes.  Operating parameters are created by setting budgets for systems and functions.

Proposed regulations are already making some power producers nervous about next winter. More intermittent power sources are going to make the power grid a less reliable partner. The Galvin Perfect Power Initiative states the reliability comes from within each node, and resilience from a node’s neighbors. The Zombie Fortress is the ideal node to participate in a smart microgrid, whether it encompasses the back-country bolt-holes, or an in-town neighborhood. Zombie fortresses are self-aware, at least so far as energy use, and ready to trade.

Don’t plan for short term inducements and temporary incentive. Design systems the self-integrate with other systems in the facility. Design systems able to negotiate with their peers for predictable load curves, effective pre-consumption, aggressive storage and full use of “excess” energy

We need systems designed for the Zombie Fortress.

Efficiency, Resilience, and Smart Energy

Far too many of the presentations at Connectivity Week last month touted building efficiency. Efficiency is important to Smart Energy, but can also work to defeat Smart Energy. Resilience is ultimately more important than efficiency for meeting the goals of Smart Energy. What energy efficiency can do, is support energy resilience.

A Smart Grid is one that can work despite a growing volatility of supply. Today’s grid already has a reduced ability to support the ever-changing aggregate consumption by the end nodes. Buildings, houses, and industry, the end nodes of the grid, will be the basis for Smart Energy.

So far, today’s efficiency efforts have wrung the slack from the system. A system without slack becomes brittle because it has a smaller margin for error. The most efficient buildings are limited in how they can trim load when asked. The overall grid has reduced margins for error. An exclusive focus on efficiency drives the impulse to direct load control in the end nodes by the central systems of the energy supplier.

Resiliency is the capacity of a system to absorb disturbance and still retain essentially the same function, structure, identity, and feedbacks. At the local level, resilience is dependent on the ability to adapt and to use diverse resources to achieve the same ends. At the broader level, resilient systems are characterized by diverse participants with non-uniform responses. Homogenous collections of systems respond to a given stimulus in similar ways, resulting in “panics” or “stampedes”. Smart grids will provide many systems with a similar stimulus as power availability changes.

Smart Energy results when the end nodes are able to respond to situations announced by the Smart Grid. It is critical to note that the purposes of the end nodes are not those of the grid. The Smart Grid will present its problems with reliability and balance to the end nodes. The end nodes, whose goal is to deliver divers services to their owner / occupants will use this information to optimize their own service delivery.

Let me present two examples of systems whose proper goal is service resilience rather than energy efficiency.

Cloud computing data centers use immense amounts of power, converting it to business process and to heat. Cloud computing relies on virtual computing machines that can be started and stopped, created and destroyed as needed. Cloud data centers have a growing ability to move these virtual machines between data centers. They are using this capability to provide service resilience whether or not a given data center is operational.

Data center resilience used to be provided through physical security, redundant systems, and back-up generators. The new model provides resilience through an ability to run from the problem, moving a virtual machine from one center to the next. The cost of each data center is reduced as the redundant systems and unnecessary generators are eliminated; construction savings of more than 50% were reported. Each data center is less robust, but together the data centers gain resilience.

Resilient data centers can respond to Smart Grids by moving processes from one site to another. Cloud services are part of smart energy in ways that data centers never could be. This resilience is not built on energy efficiency; six data centers may replace one. They have achieved resilience by focusing on their own missions rather than on support of the grid.

Commercial buildings and homes can achieve resilience by focusing on the times of energy surplus. Many renewable sources on the grid are unable to find adequate markets when they are producing at their maximum. Times of energy surplus may occur every day, while energy shortages may occur a dozen times a year. When the wind is blowing, when the sun is shining, Smart Grids will let the end nodes know with low prices. It is these low prices more than peak price events that will provide the incentives for smart energy.

Periodic low prices will fund resilience in those end nodes that take advantage of them. Capturing and storing the surplus, particularly with in-process storage, makes each building better able to weather shortages. Through storage combined with efficiency, each end node will lessen the urgency to buy power now. A building that is planning around the temporary power surpluses is able to respond to shortages without loss of service. The net effect to the participant is more reliable service at a lower price than competing buildings and properties.

Over time, end-nodes that commit to on-site storage will find that their internal markets change. On-site generation will be the market for site-based energy, in preference to grid-based distribution. The better market is the internal one, wherein storage can enhance service to the building owner and occupant.

As their site-based storage grows, the technology costs will drop. With each progressive step, building resilience grows , and grid dependency is reduced. Because there are many buildings, with many owners, and many motivations, smart energy in buildings better supports the market dynamics of rapid innovation. Because the building owners are inherently diverse, and building systems naturally autonomous, building based smart energy gains resilience as a larger system of systems.

Efficiency supports this developing resilience by reducing the demands. A building that uses half as much energy need store only half as much energy. A building that uses less energy can better weather periods of limited support from grids. To the end node, the advantage of a smart grid is better situation awareness, and an improved ability to broker whatever services are needed locally for the occupants.

The largest Smart Energy opportunities are not in selling to the grid. The real opportunities are in building end-node resilience despite power whose price, quality, and availability will be more volatile. The purpose of this resilience is to better support the owner and the occupants of the end node, not to support smart grids. This focus, on the local decision maker and their needs will lead to faster adoption.