Distribution Transformer Area Multi-source Load Forecasting Method Based on LSTM + Transformer Spatiotemporal Attention Model

For the multi-source forecasting problem of distribution transformer area loads, traditional centralized regulation confronts challenges such as high communication delay and insufficient global forecasting capability. This paper presents a distribution transformer area multi-source load forecasting method based on LSTM + Transformer spatiotemporal attention model. In the forecasting phase, the cloud integrates multi-source data from smart meters, meteorological parameters, and user behavior using an LSTM + Transformer model to achieve one-day-ahead predictions. During the real-time phase, a lightweight LSTM + Attention model is deployed at the edge layer to perform ultra-short-term forecasting, establishing a dual closed-loop system of “global trend (pre-judgment) + local real-time response.” Case studies demonstrate that the model exhibits a high degree of consistency in prediction trends, enabling accurate forecasting of multi-source loads.