Grid operators no longer get to chase a single “optimal” answer. A wave of multi-objective optimisation research – from transmission topology control to microgrid dispatch – is instead handing them menus of trade-offs, and teaching the grid to choose faster than any human planner could.
Ask a control-room engineer what the “best” way to run today’s grid is, and the honest answer is: it depends which problem you’re willing to make worse. Push hardest on cost and you strain reliability margins. Chase emissions cuts and you may need more switching actions, more curtailment, or more expensive balancing. Every lever an operator pulls trades one objective against another, and as demand grows and renewable generation becomes less predictable, those trade-offs are multiplying faster than manual planning can keep up.
That’s the problem a growing body of multi-objective optimisation (MOO) research is trying to solve — not by finding the single best configuration for a grid, but by mapping out the full menu of good ones and letting the algorithm do the mapping in minutes rather than hours.
Why “optimal” is the wrong question
Classical optimisation looks for one best answer to one objective: minimise cost, say, or minimise losses. Real grids rarely offer that luxury, because the objectives that matter — cost, emissions, reliability, equipment wear, voltage stability — pull in different directions at once. MOO methods don’t try to collapse these into a single score. Instead, they search for a Pareto front: a set of solutions where no objective can be improved without making another one worse. An operator (or an automated planner) then picks the point on that front that fits the day’s priorities.
The algorithms that generate these fronts vary — evolutionary methods such as NSGA-II and NSGA-III, swarm-based methods such as particle swarm optimisation (MOPSO), mixed-integer linear programming, and increasingly reinforcement learning — but the goal is consistent: give the engineer real options, fast, at a scale that matches an actual transmission or distribution network rather than a textbook test case.
Transmission: teaching the grid to plan its own topology
The most demanding test of that scale claim is topology control on a real transmission network — deciding, hour by hour, how substations should be configured to relieve congestion without triggering unnecessary switching. Researchers from RTE, TenneT and TU Delft, presenting at the ACM e-Energy 2025 conference, built a two-phase system that combines reinforcement learning with a fast planning step to generate full day-ahead topology plans.
Tested against a full year of historical data from TenneT — the Dutch–German transmission system operator, covering a network of 1,659 substations and 1,338 transmission lines — the method produced day-ahead plans in four to seven minutes, solving 100% of in-distribution test days and more than 75% of out-of-distribution days, while averaging fewer than two topology switches per day. The authors estimate that adopting the approach could save transmission operators millions of euros a year in avoided congestion costs, simply by giving control rooms a fast, repeatable way to explore trade-offs that today are worked out largely by hand.
Distribution and microgrids: the same trade-offs, smaller footprint
Scale the problem down from a national transmission network to a distribution grid or a single microgrid, and the same tension between cost, emissions and stability reappears — just with different levers.
A 2025 study on a distributed energy system in China’s Hunan province combined multi-objective optimisation with deep reinforcement learning to balance operating cost, pollutant emissions, voltage profile and renewable consumption simultaneously. Benchmarked against a standard particle swarm approach, the method cut operating costs by 11.9% and emissions by 6.1%, while improving on comparison algorithms by 12–16% on a composite reward measure.
At the microgrid level, a separate 2025 study modelled a hybrid system of solar, wind, microturbine, diesel and battery capacity, optimising across seven objectives — grid costs, generator fuel and start-up costs, renewable costs, emissions penalties, demand-response incentives, transmission losses and voltage regulation — using a global-criterion formulation solved with mixed-integer linear programming. Teaching-learning-based optimisation outperformed particle swarm, differential evolution and biogeography-based alternatives in the comparison; layering in demand-response programmes cut operating costs by up to 3.5% and losses by up to 3.7%, and running the microgrid in standalone mode cut annual CO₂ emissions by 51.6% against a conventional grid connection.
Demand side: optimising the grid from the meter up
MOO is also moving into demand-side management, where the objectives multiply again — cost, peak demand, renewable self-consumption, user comfort, battery wear, grid stability and emissions, often across residential, commercial and industrial customers with different priorities at once. A 2026 study proposed a hybrid method — Reference-Guided MOPSO — that fuses the convergence speed of particle swarm optimisation with the reference-point mechanism from NSGA-III to keep solutions diverse across seven objectives at once. Tested across sector types, it delivered a 20% cut in operating cost, a 19.7% reduction in peak-to-average demand ratio, an 18-percentage-point increase in renewable utilisation, and a 30% improvement in both user comfort and battery-degradation metrics, alongside load and price forecasts accurate to within roughly 3.5–7.2% (MAPE).
What this means for engineers on the ground
None of these methods hands an operator a single “correct” grid configuration, and that’s the point. What they hand over instead is a short, ranked list of genuinely different ways to run the system today — this one cheaper but with more switching, that one greener but tighter on margin — computed fast enough to matter inside an operational planning cycle rather than a research timeline.
The direction of travel across transmission, distribution, microgrid and demand-side work is the same: fewer hand-tuned heuristics, more algorithms that explore the full trade-off space and hand back options an engineer can sanity-check rather than simply trust. The TenneT results in particular suggest that the barrier to adoption is no longer computational — four to seven minutes per day-ahead plan is well inside an operational window — but organisational: getting these tools validated, trusted and wired into real control-room workflows. For engineers watching this space, that’s the gap worth tracking next.
Sources:
- Efficient Multi-Objective Optimisation for Real-World Power Grid Topology Control — ACM e-Energy ’25
- Towards Efficient Multi-Objective Optimisation for Real-World Power Grid Topology Control — arXiv preprint
- Multi-objective optimization models for power load balancing in distributed energy systems — Energy Informatics, Springer Nature
- Optimizing microgrid performance: a multi-objective strategy for integrated energy management with hybrid sources and demand response — Scientific Reports
- An Optimization Framework for Intelligent Load Management Across Smart Grid Sectors Using Reference-Guided MOPSO — Electricity (MDPI)
