The Early Days of AI and Its Real-World Applications
These days, AI is everywhere. When I started MindTitan, an AI agency, 8 years ago, everyone talked about using AI, but few actually did anything, let alone anything useful. These days, I chat with a large language model daily and swipe away my phone's helpful suggestions for navigating the digital world while an AI-illustrated ad runs in the background. That's the part that I see touching my daily experience, but I know that what I directly see is just a fraction of the algorithms affecting my daily life.
Building Practical AI Solutions
As a Data Scientist, AI consultant, and Machine Learning Solutions Architect, I've built AI systems that help people collect and process information, directly speak to them or quietly work in the background, reaming tons of data to categorize it, detect faults on production lines or flag potential tax-evaders. There are as many use cases for AI as there are tedious tasks people have to do and then some, and while the apparent ecological impact of the current hype around LLM-s troubles me, I see the good it can do along with the dangers of misuse.
AI at GridRaven: Tackling Grid Bottlenecks with Smarter Solutions
Now, at Gridraven, I'm far from the directly visible part of AI and, maybe ironically, trying to make it work to benefit the environment. Specifically, we're trying to address the ever-increasing need for electricity and the bottlenecks in transmitting it without building new infrastructure and allowing more renewables on the grid. The way to do it is by reducing inefficiencies in the existing grid by enabling it to adapt to environmental conditions.
One of the things limiting power transmission is that the wires used for it get hot and break down - more power gets you more heat, so strict limits are in place.
However, weather can have quite a significant effect on this, both by cooling and heating the wires. Theoretically, it is substantial enough to increase the transmission limits by up to 3x in some cases. For a number of good reasons and a few bad ones, this and the changing nature of weather is largely ignored.
Balancing Risk and Flexibility in the Power Grid
There are a lot of rules and rigid standards in the energy sector to mitigate the volatility of nature and for very good reasons. Nature is a complex system that's hard to predict, and the power grid is a critical infrastructure you don't want to mess around with. Adding flexibility to a system like this, without increasing risk, requires making predictions about the environment and assessing the uncertainties about the predictions themselves. Any AI system aiming to help here should provide sufficient data for a person to make the call to allow more power through the grid, knowing they're not taking on unnecessary risks. This puts any viable solutions at the intersection of atmospheric and geoscience, machine learning, and Bayesian inference, which is surprisingly unexplored, considering how cool it sounds.
Adapting Existing Research for GridRaven's Unique Needs
As there's little to grab and use off the shelf, our work is about finding the bits and pieces from related research that show potential in related uses and adapting these to our rather specific case. Luckily, what we don't lack is data - there are decades of remote sensing data from satellites packed with sensors, weather simulations, and historical forecasts, observation data from the polar bears to penguins, etc. There's also more and more highly detailed data, mapping the earth at <1m resolution. The algorithms used for modeling the climate and weather can't even use this level of detail not only because they're not designed to take advantage of the information, but the scale of computation itself would be prohibitive.
Teaching AI to Understand Environmental Interactions
The algorithms we build are not designed to solve the same problem, which makes it both harder and easier. We don't have to simulate millions of square kilometers of atmosphere, but we do need to understand how every tree and building influences the weather a few tens of meters above ground and how the density and type of vegetation change the wind speed and direction. In direct opposition to the term, Artificial Intelligence lacks what we call common sense and starts out without a model of the world - both things you need to make deductions like "Dense forest blocks wind if directly in the path of it".
A lot of our work is teaching the machine what the earth is, what kinds of stuff grow on it, what the atmosphere is and how they influence each other in different situations. There aren't really labeled data sets around for that though, but the abundance does have a quality of its own and self-supervised learning methods allow us to have the model teach itself to understand the structure of the data and what parts of it are more important than others. It learns to extract useful patterns and understand how terrain, vegetation and weather work in general. Not exactly "common sense" as we know it, but something in that direction.
An additional problem we have in this field, though, is the inherent inaccuracy of the inputs - all measurements have some error, different sources have different errors and measurements are rarely continuous. Forests are cut down, buildings get built and weather stations act up or report measurements infrequently. We need the model to understand that this can happen so it can inform us if something "feels weird" at a specific time and place and we should trust its predictions less and try to help it a bit with better and/or more data.
This means incorporating ideas from climate and weather modeling, landslide detection, deforestation monitoring, autonomous drones, remote sensing data fusion, multimodal learning, etc., and combining it in a probabilistic neural network - an architecture that attempts to adjust its confidence to the amount of available evidence. It does mean that our AI is always a bit unsure about its output, but in this case, its exactly what we want.
Instead of saying that the wind will blow 5.398m/s tomorrow at 14:29, its predictions are more akin to "Not sure, but it's almost definitely gonna be at least 3m/s".
Add some confidence intervals to it, and that's sufficient to make informed decisions on acceptable risk while gaining flexibility.
GridRaven’s mission is to rethinking how we use our existing power grid, making it smarter and more adaptable. By applying AI to optimize transmission capacity dynamically, we’re not just addressing today’s energy demands but building the foundation for a resilient and sustainable energy system. As we continue to refine our technology, our commitment remains clear: to empower grids around the world to work harder, safer, and cleaner—unlocking their full potential.
