Customizable scenario generation for clean energy bases

In this study, we enhance the Least Squares Generative Adversarial Network (LSGAN) by integrating residual connections and self-attention mechanisms, resulting in the Res-LSGAN-Attention model, which exhibits improved stability and performance in scenario generation. To further enhance optimization efficiency, we propose a novel strategy that couples Halton quasirandom point set with the Equilibrium Optimizer (EO) and Archive Evolution Path (AEP) within the FC-MOEO/AEP framework, leading to the Halton-FC-MOEO/AEP algorithm. By applying a multi-objective optimization process to the hyperparameters of the Res-LSGAN-Attention model, we obtain 20 Pareto-optimal solutions, each highlighting different optimization trade-offs. These solutions enable the creation of customizable clean energy base models for generating wind power, direct normal irradiance (DNI), and load data. Comparative experiments against 12 widely used multi-objective optimization algorithms demonstrate that the proposed approach significantly outperforms existing methods in both Pareto frontier solution quality and optimization efficiency. The three kind of correlations between decision variables and objective functions in the optimization results are calculated and analyzed in detail. Compared with state-of-the-art scenario generation model under the same number of training rounds, the model obtained through our proposed algorithm demonstrates superior performance. This work provides a high-fidelity and customizable scenario generation framework, supporting improved decision-making in clean energy base planning and operation. • Achieves high-quality customizable scenario generation with multi-objective trade-offs. • A novel approach combining Halton quasirandom point set with fast-converging optimization. • A Robust Similarity Framework with high distributional similarity and reliable estimation.