Tehnički vjesnik, Vol. 33 No. 4, 2026.
Izvorni znanstveni članak
https://doi.org/10.17559/TV-20260113003285
Research on Optimization Method of Renewable Energy Prediction Feature Model Based on Deep Learning Model
Jiongju Hao
; School of Information Engineering, Yellow River Conservancy Technical University, Kaifeng, 475004, China
Lulu Zhao
; School of Information Engineering, Yellow River Conservancy Technical University, Kaifeng, 475004, China
Jianzhuang Li
; Yellow River Conservancy Technical University, Kaifeng, 475004, China
Hanzheng Sun
; School of Civil Engineering and Transportation Engineering, Yellow River Conservancy Technical University, Kaifeng, 475004, China
*
* Dopisni autor.
Sažetak
As the share of renewable energy in power systems continues to grow, improving prediction accuracy has become critical for enhancing system flexibility and reducing operational costs. In this paper, we propose two novel optimization methods tailored for renewable energy prediction. First, we develop an Adaptive Binary Genetic Algorithm (A-BGA) that introduces population-diversity-driven dynamic crossover and mutation rates, and reformulates the fitness as a bi-objective trade-off between prediction RMSE and feature cardinality. Through 80 stratified bootstrap replications (instead of a simple 100-run repetition) and Friedman-Nemenyi statistical testing, we identify Pareto-optimal feature subsets that significantly outperform conventional fixed-rate BGA baselines. Second, we introduce a spatio-temporal scenario generation method using a deep generative model that captures the dynamic distribution of renewable energy output in an unsupervised manner, without requiring any prior statistical assumptions. By integrating point prediction results, the model constructs a stochastic optimization framework capable of directly generating a large number of realistic future scenarios. Unlike traditional sampling techniques, the generated scenarios effectively represent the intermittency, randomness, and volatility of multi-location renewable energy generation while preserving both temporal and spatial correlations. Together, these two contributions form a comprehensive framework that enhances prediction efficiency and uncertainty modeling, offering a practical and scalable solution for high-penetration renewable energy systems.
Ključne riječi
binary genetic algorithm; conditional wasserstein-GAN with gradient penalty; generative model; feature input selection; uncertainty
Hrčak ID:
348709
URI
Datum izdavanja:
30.6.2026.
Posjeta: 0 *