Scholarlink

<p>Short-term univariate forecasting of nanogrid power consumption is essential for balancing supply and demand in small-scale energy systems. Existing benchmarks, often focused on large utilities or renewable generation, inadequately address the non-stationary, bursty patterns inherent in nanogrid loads. This paper introduces DrahiX, a novel five-year, multi-zone hourly consumption dataset annotated with regime shifts and missing-data patterns. We also present a comprehensive open-source benchmarking pipeline, featuring standardized preprocessing, rolling-window splits, diverse error metrics (point-forecast and shape-aware), and quantitative measures of forecast difficulty (e.g., permutation entropy, change-point counts). Through rigorous evaluation of statistical, machine learning, and deep learning models, our findings indicate that deep learning architectures, particularly Time2Vec-BiLSTM, demonstrate superior performance compared to simpler models, especially under conditions of higher series non-stationarity and complexity. The public release of the DrahiX dataset, all code, and hyperparameter configurations ensures full reproducibility, aiming to accelerate research and enhance the operational efficiency, renewable self-consumption, and resilience of nanogrid systems.</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

Bridging the Gap in Nanogrid Load Forecasting: The Drahi-X Dataset and a Standardized Benchmarking Framework

  • 期刊论文

  • 发表日期：

    2025年12月13日

查看全文
  • Aurélien Aurus,
  • Guillaume Guerard

摘要

Short-term univariate forecasting of nanogrid power consumption is essential for balancing supply and demand in small-scale energy systems. Existing benchmarks, often focused on large utilities or renewable generation, inadequately address the non-stationary, bursty patterns inherent in nanogrid loads. This paper introduces DrahiX, a novel five-year, multi-zone hourly consumption dataset annotated with regime shifts and missing-data patterns. We also present a comprehensive open-source benchmarking pipeline, featuring standardized preprocessing, rolling-window splits, diverse error metrics (point-forecast and shape-aware), and quantitative measures of forecast difficulty (e.g., permutation entropy, change-point counts). Through rigorous evaluation of statistical, machine learning, and deep learning models, our findings indicate that deep learning architectures, particularly Time2Vec-BiLSTM, demonstrate superior performance compared to simpler models, especially under conditions of higher series non-stationarity and complexity. The public release of the DrahiX dataset, all code, and hyperparameter configurations ensures full reproducibility, aiming to accelerate research and enhance the operational efficiency, renewable self-consumption, and resilience of nanogrid systems.