Using Machine Learning to Reduce Energy-Related Carbon Emissions from Buildings
Climate Change is currently the biggest issue that faces us as a planet and is the most real existential threat to not only the human race but to the millions of species that exist in our biosphere. Though several major political bodies have either remained in denial or have chosen to ignore/minimize the issue at hand, the effects of climate change are getting increasingly harder to ignore. Storms, droughts, floods, and fires have become stronger and a lot more frequent. The 2018 intergovernmental report on climate change estimated that the world will face catastrophic consequences unless global greenhouse gas emissions are eliminated within thirty years. Even then, carbon emissions are constantly on the rise.
Finally doing something about the climate crisis involves two steps — Mitigation, which focuses on reducing emissions, which is what we are going to discuss here, and Adaptation — i.e. preparing for the possibly dire, inevitable consequences that await us. Mitigation will require us to bring multi-faceted change to several fields, which are carbon emission contributors, and in this discussion, we are going to discuss the applications of machine learning in doing so. In the reduction of carbon emissions, our main focus is always going to be optimization, which is where machine learning comes in. In this particular discussion, our area of focus is the optimization of processes that create carbon emissions in buildings.
Buildings and their construction together account for 36 percent of global energy use and 39 percent of energy-related carbon dioxide emissions annually, according to the United Nations Environment Program. But, the CO2 emissions “pie” is a lot more complicated than a sector-by-sector breakdown, as each of these areas contributes to the others. Hence, understanding how exactly buildings contribute to carbon emissions is important for addressing the problem. It turns out, that buildings contribute to climate change on three levels: in how they are constructed, how we use them, and their location. Even the emissions from buildings can be separated into two types based on their sources — namely, operational carbon emissions and embodied carbon of a building. Operational carbon emissions are the emissions from daily energy consumption by the processes of the building, such as heating, cooling, lighting, etc. Globally, building operations account for about 28 percent of emissions annually.
The problem of GHG emissions from buildings is one of the easiest to fix. Yet, buildings contribute to almost a quarter of the total energy-related carbon emissions as of today. This problem can actually be fixed using a combination of easy to implement fixes and state of the art strategies. In fact, it is possible to bring down a building’s carbon emissions by almost 90% in this way, helping the building to consume almost no energy. Improving energy efficiency in a building will not just help protect the environment, but will also benefit the occupants, who will have to pay a minimal amount for energy, economically.
The issue that faces us in the implementation of energy optimization methods is that it is impossible to define a single unanimous solution for every building in question, as buildings themselves are highly heterogeneous — they vary according to age, construction, usage, and ownership, and hence optimal strategies vary widely depending on the context. For example, a building that consumes cleaner energy, may not require the incorporation of expensive features such as intelligent light bulbs, etc. Moreover, since buildings have a very long lifespan, they also need solutions with inertia. We should be able to make newer buildings highly energy-efficient, while also retrofitting the old ones to optimize their energy consumptions.
Energy consumption can be reduced majorly through urban planning, change in public policy, and even building management level infrastructure developments. On the building management level, ML can help select energy optimization strategies tuned for the specific building, thus attaining maximum success, while on the level of urban planning, we can use ML to fine-tune public policy changes, by gaining better insight into current GHG emissions and simulating possible impacts that specific reforms will have on future GHG emissions.
ML’s role in the context of this discussion is to help accelerate methods implemented to improve energy efficiency (thus reducing consumption) and help reduce the money spent on energy needs. ML can help carry this out in two ways — i) Modelling data on energy consumption ii) Optimizing energy consumption (smart buildings)
Modeling data on energy consumption basically involves making sense of the increasing amounts of data produced by measuring devices such as meters or home energy monitors. Working with this data will help power companies optimize building design and operation strategies, as the availability of this data helps machine learning algorithms predict/forecast energy demand given certain parameters/conditions pertaining to, and representative of the specific strategies or design. In the pre-machine learning era, demand forecasting was a strenuous process, requiring physical and domain-related knowledge, involving complex thermodynamic calculations. Today, ML algorithms have sped up these calculations to a great extent by either wholly ignoring the physical knowledge about the building, incorporating it into the computations, or learning to make approximations about the physical model (surrogate models) hence avoiding the need for expensive simulations.
For instance, Mocanu et al modeled building load profiles with reinforcement learning and deep belief networks using data on commercial and residential buildings; they then used approximate reinforcement learning and transfer learning to make predictions about new buildings, enabling the transfer of knowledge from commercial to residential buildings, and from gas- to power-heated buildings.
We must also try and identify which appliances drive the energy consumption in the building, in a process called Energy Disaggregation. This will then help in focusing the energy efficiency and optimization where it really has an impact.
To verify the success or failure of energy efficiency interventions, statistical ML offers methods for causal inference. An example of this is the use of Lasso regression on hourly electricity consumption data from schools in California which helped understand that the implemented energy efficiency measures did not actually achieve the estimated level of savings. This is an instance of counterfactual predictions, and the use of deep learning in the same.
The application of machine learning is probably most relevant to smart buildings. With an intelligent control system in place, we can reduce carbon emissions by using ML to enable the system to adapt to the general usage trends of the residents, respond to signals from the grid and hence reduce costs for the consumer, by providing flexibility to the grid. We will also be able to integrate lower-carbon electricity sources.
It is quite apparent that several systems that consume energy can be a lot more efficient, thus reducing energy consumption and hence carbon emissions, and most of these systems are highly critical and hence huge consumers of energy. This requirement for increased efficiency is wide, ranging across all kinds of devices.
A good example of this is the infamously energy heavy, highly inefficient HVAC systems. In fact, these systems account for more than half the energy consumed in buildings. There are multiple ways to increase efficiency and reduce energy consumption in HVAC systems most of which can be achieved with the use of ML, the main three of which are — forecasting temperatures needed across the systems, improved control with the aim of achieving those temperatures, and fault detection. As mentioned before, modeling energy use of a building, previously done with thermodynamic calculation heavy methods, can now be computed easily, and for a much lower cost using deep belief networks. For control, Kazmi et al. used deep reinforcement learning to achieve a scalable 20% reduction of energy while requiring only three sensors: air temperature, water temperature, and energy use.
ML can also be used to automate building diagnostics and maintenance using fault detection, a very important component of increasing energy efficiency, as it is often faulty systems like refrigerant lacking refrigerators. We can also use ML to automate lighting and heating based mainly on occupancy and occupancy-based patterns, which not only reduces energy consumption but also helps augment occupant comfort. ML can help these systems firstly detect occupancy, recognize patterns in occupancy and finally dynamically adapt to changes in occupancy patterns using deep neural networks, using input from Wifi Signals, occupancy sensors, UV Sensors, general appliance power consumption data, etc.
In the discussion, what has become clearly apparent is the need for automated adaptation to specific conditions, based on patterns that may or may not be easily observable/traceable to humans, and in most cases cannot be controlled by us, which is where machine learning proves to be useful. The discussion indicated that very often, adaptation is the key to efficiency, and efficiency the key to reduction, here the reduction of energy consumption thereby reducing emissions of GHGs. Then again, even though machine learning might seem like the perfect solution in several cases relevant to this discussion, we must not forget that the components that consist a smart building consume energy too, and are often quite expensive, which might not be feasible when it comes to a broader application. Secondly, rebound effects are likely and can cause almost a 10–20% additional energy consumption in the building. Moreover, when these systems are built with the aim of cost reduction, they don’t necessarily help reduce GHG emissions or increase energy efficiency. Moreover, the construction of the components required in such smart buildings often causes the depletion of mineral resources, some of which are extracted using processes that cause widespread deforestation and have several other negative environmental effects. But in the hope that with improving technology such issues can still be battles, and by weighing the pros of such systems over their cons, they are still quite revolutionary.