Scholarlink

This study presents a structured, progressive training approach to systematically quantify and improve uncertainty estimation in hourly building energy forecasting, leveraging a machine learning approach. Starting with deterministic LightGBM modeling and advancing through single-model Quantile LightGBM to ensemble-based Quantile LightGBM, the research explicitly captures both aleatoric (data-driven) and epistemic (model-driven) uncertainties. Deterministic and median Quantile LightGBM models exceeded the ASHRAE-14 guideline, validating the accuracy and reliability of quantile regression. Single-seed Quantile LightGBM effectively quantified aleatoric uncertainty through prediction intervals at 80 and 95%. Ensemble modeling further improved prediction interval coverage at the 80% nominal level, increasing the overall Prediction Interval Coverage Probability (PICP) from 68.84 to 72.60% and reducing the Absolute Coverage Error (ACE) from 11.16 to 7.40%, thus enhancing predictive reliability by achieving coverage closer to the target level. Future research directions include expanding the ensemble size to 30 seeds, enhancing statistical robustness, and exploring model interpretability through feature importance and SHAP values. Differences in performance between workdays and weekends/holidays will be investigated in subsequent journal publications.

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

Progressive Uncertainty Quantification From Deterministic to Ensemble Probabilistic Forecasting for Hourly Building Energy Predictions

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Kornkamon Tantiwanit,
  • Ploy N. Pratanwanich,
  • Apiparn Borisuit,
  • Phanchalath Suriyothin

摘要

This study presents a structured, progressive training approach to systematically quantify and improve uncertainty estimation in hourly building energy forecasting, leveraging a machine learning approach. Starting with deterministic LightGBM modeling and advancing through single-model Quantile LightGBM to ensemble-based Quantile LightGBM, the research explicitly captures both aleatoric (data-driven) and epistemic (model-driven) uncertainties. Deterministic and median Quantile LightGBM models exceeded the ASHRAE-14 guideline, validating the accuracy and reliability of quantile regression. Single-seed Quantile LightGBM effectively quantified aleatoric uncertainty through prediction intervals at 80 and 95%. Ensemble modeling further improved prediction interval coverage at the 80% nominal level, increasing the overall Prediction Interval Coverage Probability (PICP) from 68.84 to 72.60% and reducing the Absolute Coverage Error (ACE) from 11.16 to 7.40%, thus enhancing predictive reliability by achieving coverage closer to the target level. Future research directions include expanding the ensemble size to 30 seeds, enhancing statistical robustness, and exploring model interpretability through feature importance and SHAP values. Differences in performance between workdays and weekends/holidays will be investigated in subsequent journal publications.