Short-term load forecasting (STLF) is crucial for the economic operation and planning of power grids. The modern power system is becoming increasingly complex and diversified on both the source and user sides. Facing the challenge of extracting features from load series, this paper proposes an STLF method based on load series imaging and SE-ResNet. This paper first encodes load series into a 2D Gramian Angular Field (GAF) representation. Then, the 1D ResNet-50 and 2D ResNet-50 are utilized to adaptively extract features from 1D load series and 2D load images in GAF representations. The output of the final fully connected layer of both the 1D ResNet-50 and 2D ResNet-50 corresponds to the predicted load values. Finally, the 2D ResNet-50 with an added Squeeze-and-Excitation (SE) channel attention mechanism is employed for STLF. The experimental results demonstrate that: (1) compared to the predictions from the 1D ResNet-50, the 2D ResNet-50 demonstrates superior prediction performance, and (2) the 2D SE-ResNet-50 achieves the best load forecasting results.