More capacity for day-ahead energy markets.
I came away from this year's conference with the feeling that there has been a significant shift in the specialised world of Dynamic Line Ratings (DLR). A year ago, it seemed that DLR was equated with hardware, but now, everyone seems to be aware of the concept of sensorless DLR. A year ago, it felt like we were explaining the concept for the first time, but now people were keen to ask: "Why sensorless?" and "How do you know it works."
Answering "why sensorless" is straightforward. The majority of energy market decisions are made at least one day ahead. Where there are no markets, dispatch decisions are also usually taken at least one day in advance. Therefore, most of the possible economic benefits from DLR are also obtained if it is available at the same time, allowing for the lowest-cost plants to run even if they're further away. This helps utilities make more money while improving consumer access to low-cost energy.
There are no sensors that can measure what will happen tomorrow. Numerical weather predictions are used in forecasting by all DLR solutions that I'm aware of. In some solutions, sensors are used to improve forecast accuracy by leveraging the measured historical time series and learning the error patterns in past forecasts. Gridraven has shown that it is not necessary to install sensors to have accurate forecasts, unlocking sensorless DLR for entire grids.
Applying machine learning to improve DLR forecasts
I like to think of machine learning (ML) as an extremely sophisticated search function. Our goal is to determine the conductor cooling conditions in each span. The presence of trees and hills will slow down the wind in particular, so we need to account for the landscape.
The classical approach would be to build a deterministic physical model that simulates the flow of wind around obstacles. But this is extremely computationally expensive and cannot scale up globally to cover all power lines in the world with hourly updates.
Instead, we use ML to "search" through the past measurement histories of tens of thousands of weather stations for situations in which the trees and hills surrounding a station, as well as the weather conditions, looked "similar" to the one we're predicting for. In this way we can accurately determine the wind conditions and hence DLR in this span.
A further benefit of the ML approach is that the machine can also "search" for the statistics of the confidence level of each forecast based on similar situations in the past. Based on billions of individual measurements, the neural network can accurately determine the probability for each prediction to be correct. Confidence intervals make the DLR predictions usable in the control room.
But how do we know it works?
At CEATI, during his 45-minute presentation, our CTO, Dr Henri Manninen presented exactly 17 graphs of measurements from a 110 kV overhead line in Estonia from our pilot with the Transmission System Operator, Elering. This overhead line is in a heavily forested area and we presented how our predictions compare with measurements.
The model had not seen the measurements from this line before, but was able to predict wind conditions with 60% higher accuracy than the best available numerical weather prediction while maintaining the 95% confidence interval 95.1% of the time.
The resulting dynamic line ratings were 40% higher than the seasonal rating. Notably, ambient adjusted ratings were 30% higher than the seasonal ratings, so the additional gain from accounting with the wind at the 95% confidence interval was an additional 10% of capacity.
Judging by the responses to the FERC advanced notice on proposed rulemaking on DLR, one of the main barriers to DLR adoption is the high cost of the hardware. Sensorless DLR does not have this barrier. Once line rating teams have completed the implementation of ambient adjusted ratings, sensorless DLR can simply be switched on.
To balance out the overseas travels and indoor days, Henri and I found time for trail running around the conference in the beautiful hills in the Joshua Tree National Park. Henri has completed a full ironman in under nine and a half hours, but thankfully, that was a few years ago, and he is currently not training for another one, so I was able to keep up!
We are certainly going to return to CEATI. The conferences are very professionally organized, and the quality of lunchtime conversations is excellent.
