Machine Learning is Becoming Increasingly Important to Keep Your Homes Lit
Machine learning is beyond image classification and speech recognition. Know how it is becoming crucial to power our homes.
Have you ever wondered how is it that whenever you switch on your lights, they work? How do the generation stations know that you are going to press the switch in 3..2..1… Well, they don’t! Whenever you press a switch to light up your room or put on your air-conditioner, you send oscillations across the complete power grid. These oscillations may be insignificant for one light bulb, but when you consider a whole building being lit up together, it can send waves across the complete power grid of the town — and we as power engineers, really fear that. And what happens if those waves don’t stop?
What’s the problem?
As a kid, this problem seemed a more ghostly problem for me — as I would run to my mom and dad to save myself from the imaginary ghost in the dark. But as I grew up enough to do a PhD in this field, I realized it is much bigger than that. A problem which can cost a country like the US, $150 billion annually. But what do you think is the problem? — it’s the complexity of such a huge power system. Don’t believe me, check the image below.
Can you count those lines? — surely not. This represents the US interconnected power grid. Now think again — if you suddenly light a large building, how will such a complex network of transmission lines and generating stations know and deal with the waves, that I told you earlier about, and meet your power demand?
Traditional control methods have proven successful in controlling power grids when they were much smaller, simpler and not as interconnected as they are now. However, with a more complex system, distributed private generation such as solar and wind, uncertainties prove to be a risk to such traditional techniques of stabilizing a power grid.
Here is where Machine Learning comes in!
I am very certain, whenever you hear ML, a few thoughts come to your mind.
This is for computer scientists.
Big companies like Google and Facebook use this.
It’s used for speech and image recognition. And maybe predict stocks value.
Don’t worry, I was wearing the same hat a couple of years ago. I could not think of any other thing when I heard ML and neural networks. I confess, I didn’t understand it either, but whoever knew it, they seemed magical gods to me. But I am sure even God didn’t know that one day I would be doing a PhD in this field, and would love doing it.
From here on, it may become a bit technical, but I promise to keep it comprehensible for everyone and to inspire a few bright minds for this field of research.
State of the art ML research for power systems
The modern electric utility industry in America began with Thomas Edison’s invention of the first practical light bulb in 1879. Soon enough, the first centralized fossil power plant was built and put into commission at New York. Since then, a lot has changed. Companies moved from direct current (DC) to alternating current (AC). More homes, more factories became part of the networked system. Till about two decades ago, the power grid remained predominantly powered by fossil fuel. However, recent advancements in renewable technologies and other alternative sources for power generation has made the grid ever more complex and also opened the field for amazing and novel research work. I will try to highlight a few predominant areas where I find the most exciting ML research is currently happening — but mind it, I assure you there is a lot more!
Wide-area monitoring systems (WAMS)
WAMS consists of sensors, servers, and algorithms which act over a wide-area, like the complete North American power grid to maintain system stability so that you and I, can continue operating our air-conditioners in hot weather. In the last decade, a lot of fast sampling sensors called Phasor Measurement Units (PMUs) were installed which enabled real-time monitoring and control of the whole system.
The sheer size of the grid, huge amounts of data flow and widespread uncertainties in the system arising from varying power consumption and penetration of solar and wind-power poses a challenge for traditional control to stabilize such systems. Therefore researchers are moving towards ML based control and monitoring of such gigantic systems.
Enormous work is being done to use the Universal Function Approximation (UFA) property of neural networks to approximate the unknown dynamics of the power system using huge amounts of data being received from the PMUs. This will not only enable better insight into the properties of the system, but also assists in making real-time and more informed control decisions. This problem does not only interests control theorists, but even big data scientists and power engineers need to work hand in hand to solve this mammoth problem.
Take the closed loop control of a system shown in the figure. We want to track the input reference signal, however, due to uncertainties in the system, traditional controllers fail to provide desirable results. On the other hand, a neuro-controllers, which basically takes the error signal (e) as the input and outputs an actuation signal to the system. The learning block evaluates the performance of this neuro-controller by evaluating the tracking signal using a desirable performance metric and tunes the parameters of the neuro-controllers accordingly. This is also called “reinforcement learning” in the controls community, which is another very popular research being adopted widely by power companies.
Not only specific to NNs, but, in general, some amazing research is also going on in using Reinforcement Learning (RL) — which is an application of ML and system identification techniques to solve this challenging problem. If you never knew these terms are also part of ML, congrats! — you just got to know how diverse ML is.
Load forecasting
Forecasting electricity consumption is of supreme importance for effective energy demand management. Poor forecasting and incorrect decision making can not only cost millions of dollars for power generating utilities, but can also leave consumers for hours without electricity causing loss in business and productivity.
Traditional methods of load forecasting relied on time series analysis and regression models. These models have given desirable results in the past, but their trustworthiness and reliability are becoming questionable in recent times. Though not solely responsible, much of it is attributed to the influx of huge amounts of renewable power in the power grid.
The power generation by such renewable energy sources are not only uncertain, but also very dependent on the short and long term weather forecasts. To make it even more cumbersome, factors such as weather, national holidays, the recent boom of electrical vehicles (EVs) etc. make it even more challenging to make accurate forecasts. Traditional methods perform poorly when forced to consider such a huge number of parameters to make forecasts.
In the past decade, a lot of researchers have taken interest in solving this complex problem using complex ML structures such as deep learning using neural networks. The figure by Moon et al. shows that how a neural network can be used to make a load forecast using variety of information as input such as the date, weather and even holidays. The problem of accurate load forecasting is so fascinating and crucial for power engineers, that researchers have studied a wide variety of ML techniques, not limited to just deep learning for predictions. For example, a popular neural network structure, known as Long Short Term Memory (LSTM) framework has been studied tremendously to process time-series information to make accurate short-term load forecasts which become important in day to day planning of power companies.
Energy markets
You may or may not have heard of this term. Essentially, energy markets are markets that deal with trade and supply of energy or electricity. The bid for energy plays a major role in deciding profits for power utilities. This market relies heavily on accurate load forecasting. Generally speaking, most markets have a day ahead pricing and real time pricing. It depends on varieties of factors such as expected cost of energy, congestion and transmission losses, reliability etc. The operator, based on this knowledge and the bids he gets, decides on buying and selling energy and the right price. However, the dynamics of this pricing are highly non-linear and very difficult for the operators to decide the “optimal” pricing in real time.
Operators rely on algorithms to make this decision for them. Machine learning, especially a deep learning framework known as Deep Learning Extreme Value Theory (DL-EVT) is being used extensively to find such optimal solutions of non-linear dynamics for the operators to make a decision of the best price to buy or sell electricity.
Smart Grids
You might have definitely heard this term once if ever you read something about electricity and recent research. Smart grids are basically a group of sensors and actuators distributed across the grid which allow the flow of information to and fro between the consumer and the power utility. Such an exchange of information allows for automated real-time control of the grid making it highly reliable and stable. The ability to make decisions online and in real-time highly relies on machine learning architectures to make the optimal decisions to minimize cost constraints such as transmission losses, stability constraints and even customer comfort.
What? — customer comfort? How ?
Some utilities control temperatures of large buildings by effectively controlling their air-conditioning units remotely through sensors and actuators. This gives the utility the complete leverage of controlling certain part of the power grid in real-time to ensure that power demand and stability requirements can be met over the complete system, while the building owner sees some great discounts on their electricity bills for letting control to the utility. Smart grids have become an important aspect of the power community, with top researchers and industrialists calling it the future of electricity. With the influx of renewables, distributed energy resources such as battery storage, smart grids are the way forward to maintain resiliency and performance. And guess what, it’s a hot area of research now, especially for those interested in Big Data and Machine Learning.
A future with smart grids is incomplete without Big Data scientists and ML engineers.
I hope you liked this article and it gave you insights on how machine learning is being extensively used for our power grids. If it excited you, that’s great. If it left you confused, that’s even better — because honestly, even after 2 years of doing research in this field, I still don’t understand a lot of it! If you would like to connect with me or ask me anything, email me at jishnudeep.kar@gmail.com or DM me on Instagram.
If you liked this, you may also like these.
