Optimizing smart meter data acquisition for residential load forecasting: Minimum requirements for interval, precision, and historical duration

초록

As Advanced Metering Infrastructure (AMI) becomes a global standard, the lack of standardized guidelines for data collection strategies remains a critical challenge for utilities. This paper investigates how smart meter data characteristics—measurement interval, precision, and historical data length—affect residential load forecasting accuracy. Using real-world smart meter data from 3000 households, 8 forecasting models, ranging from conventional machine learning to advanced neural networks, are evaluated across very short-term, short-term, and medium-term horizons. The results indicate that a 15–30 min measurement interval, at least one decimal precision, and a minimum of 6 months of historical data achieve an effective trade-off between forecasting accuracy and data storage efficiency. To support region-specific tariff schemes and the reliable selection of participants for demand response programs, a cluster-based forecasting framework incorporating Hierarchical Clustering with Dynamic Time Warping is further investigated, demonstrating that the identified optimal smart meter settings preserve comparable forecasting performance. In addition, a pre-trained transformer-based model (TimeGPT) combined with time-series decomposition achieves comparable accuracy using only 4 months of data. These findings provide practical guidance for designing data efficient AMI systems. © 2026 Elsevier Ltd.

키워드