2025-06-06 東京科学大学
図1. 開発したグループエンコーディングの概要
＜関連情報＞
- https://www.isct.ac.jp/ja/news/crsz24lgjn0a#top
- https://www.isct.ac.jp/plugins/cms/component_download_file.php?type=2&pageId=&contentsId=1&contentsDataId=1687&prevId=&key=2e5019c2940e6d17ff639ea3d02a912e.pdf
- https://www.sciencedirect.com/science/article/abs/pii/S0306261925007196
分散型エネルギーシステムにおける電力需要予測のための高次元カテゴリカルデータの新規符号化手法 A novel encoding method for high-dimensional categorical data for electricity demand forecasting in distributed energy systems
HyoJae Lee, Keisuke Kameda, Sergei Manzhos, Manabu Ihara
Applied Energy Available online: 5 May 2025
DOI:https://doi.org/10.1016/j.apenergy.2025.125989
Highlights
- The original encoding method, “Group encoding” using On/Off data in the building was proposed.
- “Group encoding” can be applied using the data simply obtained in BEMS without additional sensors.
- Electricity demand prediction using “Group encoding” achieves top level of prediction accuracy.
Abstract
There has been a gradual shift towards distributed energy systems (DES) with variable renewable energy as the main power source. However, DES present difficulties of controlling the supply-demand balance due to frequent fluctuations in electricity demand caused by human activities in a small-scale system. While energy big data can drive accurate forecasting, it is challenging in the general case due to the increasing cost of the sensor installation and decreasing density of the data in a high-dimensional feature space. To address these problems, we have studied an accurate electricity demand forecasting method without additional sensor installation cost by proposing an encoding process named “Group Encoding” (GE). The GE process was applied to existing high-dimensional binary data of On/Off state, which is already collected in building energy management systems (BEMS). GE can also solve the problem of decreasing data density in a high-dimensional feature space by reducing the dimension without loss of critical information. The effectiveness of GE was evaluated by dealing with high time-resolution data which can be used for various forecast horizons. The forecasting performance improved by 74 % in terms of MAE when using GE, compared to a typical label encoding, for a 1-min ahead forecast. The top-level of electricity forecast accuracy, 3.27 % MAPE (mean absolute percentage error) in 60-min ahead forecasting, compared to the other forecasts for single building was achieved by using only GE processed data and electricity demand.
