Resistance to hype
Engineering pragmatism
A taste for debate
In Novosibirsk, winter lasts half the year, and when your power cuts out at -35°C, you learn the real price of a reliable grid. But today, that very reliability is under threat – and the culprit is... sustainability.
It's the paradox of our time: the more homes that install solar panels and energy storage, the harder it becomes to keep the electrical grid stable. Every solar panel on a roof is a tiny power station that can both draw energy from the grid and feed it back. And when thousands of these «power stations» pop up, the grid starts to have a mind of its own.
A Problem Hidden from Consumers
Picture this: on a sunny day, hundreds of homes in a neighborhood start feeding surplus energy into the grid all at once. The voltage spikes, transformers overheat, and circuit breakers trip. In the evening, the situation flips entirely – everyone turns on their electric stoves and air conditioners, creating a massive peak load.
The traditional solution is a blunt one: just set limits. The utility company imposes strict caps: don't feed more than 3 kW back into the grid, don't draw more than 5 kW. Simple, but inefficient. Solar panels sit idle on a clear day, and their owners lose money on unused energy.
The modern approach is Dynamic Operating Envelopes (DOE). It's a smart system that recalculates individual limits for each home every 15-30 minutes, based on the current state of the grid. On a sunny day with low overall demand, the limit is higher. During a cloudy evening at peak load, it’s much stricter.
A Mathematical Nightmare for Grid Operators
It sounds logical, but there's a catch. To calculate these limits, you have to solve a system of nonlinear equations that describe the power flows in the grid. It's like trying to predict tomorrow's weather by knowing the temperature and pressure at every single point in the region – theoretically possible, but a practical nightmare.
In mathematics, there's a concept called «convex optimization» – a class of problems that can be solved quickly and with a guaranteed optimal result. And then there are «non-convex» problems, where even a supercomputer can churn for hours without any guarantee of finding the best solution.
The DOE calculation problem is a classic example of non-convex optimization. This leaves utility companies with a dilemma: either simplify the model and sacrifice accuracy, or wait for hours while a computer searches for a solution for a network with thousands of connections.
The Neural Network Solution
This is where Input Convex Neural Networks (ICNNs) come into play – a relatively new type of artificial intelligence with a peculiar but powerful property. A standard neural network can produce any kind of output for any given input. An ICNN, however, is designed so that its output function is mathematically guaranteed to remain convex.
What this means is that by training an ICNN on historical grid data, we get a model that behaves predictably, allowing us to use fast optimization algorithms.
Here's how the algorithm works:
Training Phase: The ICNN studies tens of thousands of historical snapshots of the grid's state – voltages, currents, and loads at every node. Gradually, the network learns to predict how a change in power at one home will affect the voltage at a neighbor's.
Application Phase: Instead of solving a complex system of nonlinear equations, the system uses the trained ICNN to rapidly estimate the permissible operating ranges for each home.
Practical Results
Researchers tested the method on a model of a real 123-node distribution network that included a solar farm and an electric vehicle charging station. They compared several approaches:
- No DOE: The simplest case – fixed limits for everyone.
- Classical Model: Accurate equations, but slow calculations.
- Simplified Linear Model: Fast, but inaccurate.
- ICNN Model: The new approach.
The results are impressive:
- The ICNN method completely eliminates emergency conditions on the grid ⚡
- Calculation time is cut by 80% compared to classical methods.
- Homeowners get 25-30% wider permissible power ranges.
In practical terms, this means solar panels can feed more energy into the grid at the right moments, battery storage can play a more active role in balancing the load, and the utility company can guarantee voltage stability without resorting to conservative, blanket restrictions.
Linear Relaxation: The Final Ace
To speed up the calculations even further, the researchers used a mathematical trick called «linear relaxation.» Thanks to the special architecture of the ICNN, the optimization problem can be reformulated as a linear programming problem – a class of problems that can be solved in seconds, even on a standard computer.
This is critically important for real-world energy systems, where a decision has to be made every 15 minutes for thousands of connections. A method that runs in a fraction of a second instead of tens of minutes transforms a theoretical possibility into a practical tool.
What This Changes in Practice
Implementing the ICNN approach for managing distributed energy resources opens up new possibilities:
For homeowners: More effective use of solar panels and battery storage, and additional income from selling energy during peak periods.
For utility companies: The ability to integrate more renewable sources without risking grid stability, and lower costs for backup capacity.
For the energy system as a whole: More flexible load management, a reduction in peak loads, and an increase in overall efficiency.
The Siberian Reality Check
Of course, any technology has to work not just in a lab, but in harsh reality. The ICNN algorithms demonstrated stability across various load scenarios, including extreme situations like a mass activation of electric heaters on a freezing day.
A key advantage of the method is that it doesn’t require perfect input data. Real-world grids are full of uncertainties: inaccurate consumption forecasts, fluctuating solar generation, and deviations in equipment parameters. The ICNN model trains on this «dirty» data and still produces robust solutions.
Future Prospects
This technology paves the way for creating truly smart grids, where every home becomes an active participant in the energy exchange, not just a passive consumer. In the future, we could see grids that automatically optimize energy flows in real time, making the most efficient use of every kilowatt-hour.
The next step is to integrate this with weather forecasting, consumer behavior models, and market mechanisms. Imagine a system that knows tomorrow will be sunny and adjusts today's limits to maximize the use of that free solar energy.
Mathematics conquers chaos – and there’s something reassuringly reliable about that. Especially when it’s -40°C outside, and the lights stay on thanks to smart algorithms.
