TY - JOUR
T1 - Short-term load forecasting facilitated by edge data centres
T2 - A coordinated edge-cloud approach
AU - Li, Junlong
AU - Fang, Lurui
AU - Wei, Xiangyu
AU - Fang, Mengqiu
AU - Xiang, Yue
AU - You, Peipei
AU - Zhang, Chao
AU - Gu, Chenghong
N1 - Publisher Copyright:
© 2024 The Author(s). IET Smart Grid published by John Wiley & Sons Ltd on behalf of The Institution of Engineering and Technology.
PY - 2024
Y1 - 2024
N2 - Compared to load forecasting at the national level, two challenges arise in providing accurate forecasting for LV and MV networks: (1) customers within LV and MV networks are much less, implying greater volatility within those load profiles; (2) not all customers have smart metres. Particularly, the two challenges would exacerbate forecasting performance under unexpected events, such as extreme weather and COVID-19 outbreaks. To secure accurate short-term load forecasting for LV and MV networks, this paper customised a Spatio-Temporal Edge-Cloud-coordinated (STEC) approach on a loop training structure—LV networks to MV networks to LV networks. For each LV network, this approach utilises XGboost to learn the relationship between weather data, substation-side loads, and a few accessible customer-side load data to deliver rough forecasting. Then, it adopts the rough forecasting results and accessible data for all LV networks within an MV network to train the convolutional neural networks and gated recurrent unit (CNN-GRU) network. This step provides load forecasting for MV networks and simultaneously refines load forecasting for LV networks by generating the interacting relationship between LV substations of different locations. Case studies reveal that the STEC approach successfully extrapolates the demand-varying information from long-term datasets and improves short-term forecasting performance under both normal scenarios and newly occurring unexpected scenarios for LV and MV networks. The loop training structure halves the forecasting error, compared to classic methodology by utilising the local data only for MV networks.
AB - Compared to load forecasting at the national level, two challenges arise in providing accurate forecasting for LV and MV networks: (1) customers within LV and MV networks are much less, implying greater volatility within those load profiles; (2) not all customers have smart metres. Particularly, the two challenges would exacerbate forecasting performance under unexpected events, such as extreme weather and COVID-19 outbreaks. To secure accurate short-term load forecasting for LV and MV networks, this paper customised a Spatio-Temporal Edge-Cloud-coordinated (STEC) approach on a loop training structure—LV networks to MV networks to LV networks. For each LV network, this approach utilises XGboost to learn the relationship between weather data, substation-side loads, and a few accessible customer-side load data to deliver rough forecasting. Then, it adopts the rough forecasting results and accessible data for all LV networks within an MV network to train the convolutional neural networks and gated recurrent unit (CNN-GRU) network. This step provides load forecasting for MV networks and simultaneously refines load forecasting for LV networks by generating the interacting relationship between LV substations of different locations. Case studies reveal that the STEC approach successfully extrapolates the demand-varying information from long-term datasets and improves short-term forecasting performance under both normal scenarios and newly occurring unexpected scenarios for LV and MV networks. The loop training structure halves the forecasting error, compared to classic methodology by utilising the local data only for MV networks.
KW - artificial intelligence and data analytics
KW - distribution networks
KW - load forecasting
UR - http://www.scopus.com/inward/record.url?scp=85200054490&partnerID=8YFLogxK
U2 - 10.1049/stg2.12181
DO - 10.1049/stg2.12181
M3 - Article
AN - SCOPUS:85200054490
SN - 2515-2947
JO - IET Smart Grid
JF - IET Smart Grid
ER -