Well condition prediction is crucial for enhancing oil well production efficiency and reducing fault-related losses. However, current methods are limited to predicting the condition type and cannot output the corresponding indicator diagram, which restricts practical field application. To address this issue, a short-term oil well condition prediction model based on a multi-branch feature fusion autoencoder is proposed. Firstly, to resolve the issue of weak inter-sample graphical variation trends in the indicator diagram sequence, which is susceptible to noise and similar condition patterns leading to suboptimal feature extraction, three deformable convolutional branches are designed within the encoder to focus on features along the X, Y, and XY directions of the indicator diagram, respectively. Secondly, to complement the directional features extracted by the different branches, a spatial attention-based cross-fusion structure is designed. The fused features from each branch are concatenated, and a channel attention mechanism is then applied to enhance the feature fusion effect. Finally, the feature sequence of the indicator diagram is input into a long short-term memory network to predict the features of the future condition's indicator diagram, and the indicator diagram is reconstructed based on the decoder. Simulations conducted using oil well indicator diagram sequences from actual production sites demonstrate that the proposed method achieves satisfactory performance in both oil well condition prediction and indicator diagram reconstruction.