Improved predictions for electricity distribution in time and space
Hardware grid reinforcement is good, but software grid reinforcement is preferable. Realizing the considerable potential of this asks for highly accurate five minute to five minute predictions of electricity use (and feed-in) in space and time to balance supply and demand under the precondition of grid integrity. Quite a challenge for the eight persons strong ICT for Industry workshop team supervised by Jerry Jinfeng Guo from Alliander. ‘I’m extremely impressed by the team’s motivation, performance and results’, he reacts.
‘This liaison between companies and academia is inspiring for both the researchers and for the company’
‘Six years ago, when I was a PhD student at TU Delft, NWO sent out a call for participation in ICT with Industry’, Scientific Software Engineer Guo recalls. ‘One of the cases was about optical character reading for the Royal Library at The Hague. I was intrigued and heard from colleagues who had participated before and who were excited about it. So, I decided to participate.’
Guo and his team mates from universities around, whom he hadn’t known before, were exposed to the problem on the first day. ‘We divided the tasks and went to it. In the end we managed to deliver something workable and we had learned a lot from the fast feedback we received from the supervisor. It was a really rewarding experience.’
Systematic approach
It was thus a no brainer for Guo when he found himself in the position at Alliander to bring in a case himself for ICT with Industry. ‘It is no secret that our electricity network is confronted with many challenges. One of these is to get to more efficient Newton-Raphson power flow calculation for distribution grids.’ This asks for some more explanation. The electrical grid in many regions of the Netherlands is operating very close to its capacity limits and needs to be reinforced. Matching supply with growing demand can be achieved by building more hardware such as cables and substations. This way to tackle the problem is limited by permit procedures, available space and staffing problems. It is also very costly.
A second approach is to implement software to make optimized use of the existing hardware as ‘virtual grid reinforcement’, without getting into cable overload that leads to accelerated degradation. It comes down to digital twin grid simulation. More in detail: to further enhance Power Flow (PF) calculation in the grid, for which the Newton-Raphson (NR) method is commonly used. The challenge is to rightly allocate the electricity that will be used in time and in the geography of the grid, taking into account everything from feed-in from solar farms to use patterns caused by dynamic energy contracts and, for example, football matches. The situation becomes ever more complex. But theoretically there is still room for considerable prediction optimization – at a fraction of the cost of building hardware. To be quite frank, hardware will have to be built for grid reinforcement anyway. But the software approach is not only cost effective, it also buys time given the long lead times for grid hardware building. In short: it provides energy network developers and operators such as Alliander with time to breathe.
Extremely impressed
A key challenge is to choose the right PF initial value which doesn’t lead to diversion afterwards. In practice, however, it will mostly come down to get as swiftly as possible to convergence by feeding a nearly right initial value to the NR method and staying on track after that. Guo: ‘This is normally tackled by scarce experienced engineers, who we would rather deploy for other tasks. What we need is a systematic approach to replace and possibly improve on their empirical input. To minimize iterations and to avoid divergence we envisioned three approaches beforehand. The first involves an analytical method to estimate the basin of attraction using mathematical bounds on voltages. The second is a data-driven model using supervised learning and Physics-Informed Neural Networks (PINNs). Third comes a Reinforcement Learning (RL) approach. It incrementally adjusts voltages to speed up convergence.’
It was the workshop research team’s task to make the best of a week’s time to first select the most effective of the three approaches and to subsequently develop the outlines for a solution with that approach. Guo: ‘As case owner, I’m extremely impressed by the team’s motivation, performance and results, I saw swift self-organization and efficient collaboration which led to convincing results. At present, we are really looking forward to the outcomes of the follow-ups that we plan with part of the same team.’
Quantum computing
It turned out that all three approaches hold promises, but that Reinforcement Learning, somewhat surprisingly, is the most applicable. Guo: ‘The explanation for that is the extreme complexity of the challenge. That makes it very hard for an algorithm to decide between various possible symmetrical options towards a solution. RL uses agent technology interacting with the environment to explore the problem space beforehand. It divides the problem into discrete spaces and runs preliminary PF calculations, thus creating a path towards a good calculation starting point. This preliminary step limits the number of PF calculations to make optimal use of the calculation capacity.’
Calculation capacity is no trivial point when it is about millions of nodes within a network, with lots of possible values per node. Guo: ‘The Dutch electricity grid is well integrated into European grids. That makes it resilient, but it also makes the complications much more complex. We follow the RL approach for eventual use with Q algorithms to prepare for use in future quantum computing hardware and its almost unlimited calculation power. Thus we hope to be able to apply it within a couple of years. But as RL is nicely scalable, we can already make use of it to already take the first steps in our RL roadmap.’
Better society
‘My move from academia was motivated by the wish to work on a practical contribution to society’, says Guo. ‘This workshop tackles a technical challenge, but solving this problem also leads to a better, more resilient society in which the light stays on. This liaison between companies and academia is inspiring for both the researchers and for the company. I would absolutely recommend other companies to bring in cases for ICT with Industry – as I do to my colleagues within Alliander.’
The technological revolution proceeds at an unprecedented pace, Guo argues: ‘In that situation it is in general always stimulating to look at the challenges you have with the fresh and different perspectives from outsiders. In this case, the more theoretical perspective of the researchers could provide us with practically applicable solutions as well, although we need more time for further development with the team. I would say, prepare a case, bring it to the ICT with Industry workshop and you would be surprised of what it could bring you.’
Text: Leendert van der Ent
Photo: Jerry Guo by Sjoerd van de Hucht