Annual Peak Load Forecasting for Distribution Feeders Using AMI Data and LSTM

Abstract

Accurate peak load forecasting is critical for the reliable operation and strategic planning of distribution systems. This paper proposes a Long Short-Term Memory (LSTM)-based framework that integrates advanced metering infrastructure (AMI) data with key weather indicators—including heating degree days, cooling degree days, extreme hot days, extreme cold days, persistent hot days, and persistent cold days—for annual peak load forecasting. Unlike conventional approaches, the proposed method provides enhanced spatial and temporal granularity, enabling peak forecasts not only at the system level but also at individual feeders and fine time resolutions. To mitigate systematic bias, historical forecasting errors are systematically analyzed and incorporated to model recalibration. The proposed framework is evaluated at both system and feeder levels, achieving forecasting errors consistently within 4%, which significantly outperforms conventional approaches.