This paper presents an innovative framework, based on Autoformer architecture (referred to as “load Autoformer”), designed for short-term load forecasting–a critical aspect of power system operations and energy trading. The framework introduces an enhanced autocorrelation attention mechanism, capable of capturing finer details within load data. Furthermore, an extra MLP layer is incorporated into the feedforward stage to enhance the extraction of depth information efficiently. We perform day-ahead forecasting experiments by testing the model against the Handan load dataset. Comparative analysis with various existing methods reveals the strong performance of this approach.