Abstract
Multi-objective optimization problems (MOPs) are commonly confronted in various fields, such as condition monitoring for renewable energy systems, and ratio error estimation of voltage transformers. With the increase in decision variables of MOPs, their exponentially growing search spaces are challenging for existing evolutionary algorithms. To handle this challenge, this paper suggests a coarse-to-fine large-scale evolutionary multi-objective search, called CF-LEMO. In the coarse search phase, CF-LEMO performs evolutionary search on both the original and transformed large-scale MOPs alternately, accelerating the population to approach the Pareto-optimal fronts. In addition, to alleviate the issue of diversity loss, we design a diversity preservation mechanism to preserve a well-distributed archive to support subsequent fine search. In the fine search stage, CF-LEMO conducts local search on the current population to mine high-quality solutions, which are used to update the population and archive. Then, based on the archive, the multi-objective optimization based on decomposition is employed to evolve all decision variables, so as to obtain a population with good convergence and diversity near the Pareto-optimal fronts. To assess the effectiveness of the proposed CF-LEMO, we compare its performance against four representative baseline algorithms on a benchmarks suite LSMOP1-LSMOP9 with 2 and 3 objectives. The empirical results confirm its super performance by significantly outperforming all the four competitors on 12 out of 18 benchmarks. Moreover, the experiments demonstrate the superior performance of CF-LEMO in solving multi-objective ratio error estimation problems with up to 6,000 decision variables.
Introduction
Many real-world engineering tasks are Multi-objective Optimization Problems (MOPs), requiring a balance between conflicting goals—such as maximizing system performance while minimizing cost. As these problems scale to hundreds or thousands of decision variables (Large-Scale MOPs or LSMOPs), the search space expands exponentially. This "curse of dimensionality" makes it difficult for standard evolutionary algorithms to find the Pareto-optimal front—the set of optimal trade-offs.
A critical application of this is the ratio error estimation of voltage transformers in power grids. This task involves thousands of variables and complex, non-linear relationships. To address these challenges, researchers have developed CF-LEMO, a new evolutionary algorithm that employs a "coarse-to-fine" strategy to efficiently navigate massive search spaces.
Methodology: The CF-LEMO Framework
The proposed CF-LEMO operates in two distinct phases to balance exploration and exploitation:
1. Coarse Search Phase
Instead of tackling the high-dimensional problem directly, CF-LEMO uses problem transformation. By grouping variables and assigning weights, it reduces the complexity of the search space. The algorithm alternates between searching this simplified space and the original space, allowing it to quickly "zoom in" on promising regions. A key component here is the diversity preservation mechanism—an archive that ensures candidate solutions remain well-distributed, preventing premature convergence.
2. Fine Search Phase
Once the general neighborhood of optimal solutions is identified, the algorithm switches to a fine-grained mode. It utilizes local search to perturb variables individually and polish the solutions. Finally, a decomposition-based approach is used to ensure the final population is both high-quality and diverse.
Experimental Evaluation
The algorithm was rigorously tested against four state-of-the-art competitors (WOF, NSLS, CCGDE3, and NSGA-III) using two primary datasets:
- LSMOP1–LSMOP9: A standard benchmark suite for large-scale optimization (200–300 variables).
- TREE1–TREE5: Real-world-inspired cases for Time-varying Ratio Error Estimation in voltage transformers, involving 3,000 to 6,000 decision variables.
Key Results
CF-LEMO demonstrated superior performance across both benchmarks and real-world applications. On the LSMOP suite, it significantly outperformed all four competitors on 12 out of 18 benchmarks. In the TREE tests, it maintained robustness even at 6,000 variables where some baseline algorithms failed entirely, returning non-feasible results.
Conclusion
CF-LEMO offers a powerful solution for large-scale multi-objective optimization by combining dimensionality reduction with targeted local refinement. While highly effective, future work could involve developing problem-specific transformation strategies to further enhance performance in specialized domains like renewable energy systems. Additionally, extending the framework to handle dynamic environments remains an important area for future research.
Summary updated