Decision-making problems from buying a house or a car to minimizing portfolio risk or maximizing profit in electricity markets can be modelled as optimization problems. Many real world objectives are nonconvex, noisy, constrained, or expensive to evaluate, making classical analytical methods difficult to apply. This talk introduces population-based heuristic (metaheuristic) optimization, a family of derivative-free numerical methods that search for good solutions by evolving a population of candidate points. Starting from randomly generated candidates within predefined bounds, these methods repeatedly evaluate the objective function and update candidates using operations such as perturbation, crossover, mutation, and selection until a stopping criterion is met.
Beyond classical metaheuristics, the talk will discuss how machine learning can enhance optimization workflows. ML-based surrogate models can reduce computation by approximating expensive objectives (useful when each evaluation requires simulation).
In smart-grids, these methods are especially useful because many tasks are nonconvex, constrained, mixed-integer, and simulation-heavy. Population-based heuristics have been applied to problems such as distribution network reconfiguration, optimal power flow (OPF) variants, DER and storage sizing/placement, battery and PV scheduling, Volt/VAR control, demand response, and EV charging coordination. The talk will conclude with practical guidelines on algorithm selection, constraint handling, and validation to support reliable decision-making in real applications. |