Scholarlink

<p>Deep learning (DL) models, despite their widespread adoption, present challenges related to extensive training times and a lack of interpretability. In contrast, fuzzy inference systems (FIS) offer a compelling balance between accuracy and transparency. This paper addresses the limitations of traditional Takagi–Sugeno–Kang (TSK) fuzzy models by proposing a novel Mamdani-based regressor derived from the recently introduced New Takagi–Sugeno–Kang (NTSK) framework. The resulting models are data-driven, enabling the number of fuzzy rules to be specified to manage the trade-off between accuracy and interpretability. To address the complexities of high-dimensional data, this work integrates both wrapper and ensemble techniques. A genetic algorithm (GA) is employed as a wrapper for feature selection, yielding genetically optimized model variants. Furthermore, ensemble models, namely the Random New Mamdani Regressor (R-NMR), Random New Takagi–Sugeno–Kang (R-NTSK), and Random Forest New Takagi–Sugeno–Kang (RF-NTSK), are introduced to improve robustness. The proposed methods are validated using photovoltaic (PV) energy forecasting datasets, a critical application area characterized by the intermittent nature of solar power. The fuzzy models, particularly those enhanced by genetic algorithms and ensembles, demonstrate highly competitive accuracy. For instance, the R-NTSK model achieved the lowest mean absolute percentage error (MAPE) among 29 competing methods on one of the evaluated datasets. Statistical analysis confirmed this robustness, revealing that some variants attained competitive performance with as few as a single fuzzy rule. These models frequently outperform both traditional machine learning and DL approaches while offering a simpler, more interpretable rule-based structure. To promote reproducibility, the models are provided in an open-source library named nfisis (<a href="https://pypi.org/project/nfisis/">https://pypi.org/project/nfisis/</a>), and the datasets used in this study are publicly available at <a href="https://doi.org/10.5281/zenodo.17179229">https://doi.org/10.5281/zenodo.17179229</a>.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

NFISiS: New Perspectives on Fuzzy Inference Systems for Renewable Energy Forecasting

  • 期刊论文

  • 发表日期：

    2025年12月18日

查看全文
  • Kaike Sa Teles Rocha Alves,
  • Eduardo Pestana de Aguiar

摘要

Deep learning (DL) models, despite their widespread adoption, present challenges related to extensive training times and a lack of interpretability. In contrast, fuzzy inference systems (FIS) offer a compelling balance between accuracy and transparency. This paper addresses the limitations of traditional Takagi–Sugeno–Kang (TSK) fuzzy models by proposing a novel Mamdani-based regressor derived from the recently introduced New Takagi–Sugeno–Kang (NTSK) framework. The resulting models are data-driven, enabling the number of fuzzy rules to be specified to manage the trade-off between accuracy and interpretability. To address the complexities of high-dimensional data, this work integrates both wrapper and ensemble techniques. A genetic algorithm (GA) is employed as a wrapper for feature selection, yielding genetically optimized model variants. Furthermore, ensemble models, namely the Random New Mamdani Regressor (R-NMR), Random New Takagi–Sugeno–Kang (R-NTSK), and Random Forest New Takagi–Sugeno–Kang (RF-NTSK), are introduced to improve robustness. The proposed methods are validated using photovoltaic (PV) energy forecasting datasets, a critical application area characterized by the intermittent nature of solar power. The fuzzy models, particularly those enhanced by genetic algorithms and ensembles, demonstrate highly competitive accuracy. For instance, the R-NTSK model achieved the lowest mean absolute percentage error (MAPE) among 29 competing methods on one of the evaluated datasets. Statistical analysis confirmed this robustness, revealing that some variants attained competitive performance with as few as a single fuzzy rule. These models frequently outperform both traditional machine learning and DL approaches while offering a simpler, more interpretable rule-based structure. To promote reproducibility, the models are provided in an open-source library named nfisis (https://pypi.org/project/nfisis/), and the datasets used in this study are publicly available at https://doi.org/10.5281/zenodo.17179229.