Abstract: Traditional short-term load forecasting methods do not consider factors such as holidays and users' electricity consumption habits, which result in low load forecasting accuracy. To address this issue, this paper proposes an optimized short-term load forecasting method based on multi-dimensional feature fusion and deep learning network. Firstly, the Prophet algorithm is used to extract load characteristic components from different time variables, and the fusion feature reconstruction based on the attention mechanism is carried out in combination with weather data. Then, a CNN-GRU model is used to train the fused features and obtains future short-term load forecasting values. Experimental results show that this method can effectively improve the forecasting accuracy of short-term load, which provides a support for the subsequent dispatching work of power system.