Abstract: In the workshop production process, numerous uncertainties may trigger various disturbances, which significantly affect the stability and efficiency of normal operations. To prevent and mitigate the production risks caused by such disturbances, a disturbance prediction framework tailored for digital twin workshops is proposed. Firstly, a three-layer analytical disturbance feature library is constructed based on the dimensions of "disturbance location-disturbance type-disturbance characteristics". Secondly, the multi-source heterogeneous data, including equipment operating states and production parameters, are extracted and fused to provide data support for disturbance prediction. Thirdly, a convolutional neural network gated recurrent unit with integrated attention mechanism (CNN-GRU-Attention) is designed to enhance predictive performance. Finally, the proposed method is validated through experiments on milling tool wear disturbances, demonstrating its effectiveness compared with CNN, CNN-LSTM, CNN-GAN, and CNN-GRU models. Experimental results show that the proposed method significantly improves the workshop's ability to perceive and predict disturbances, offering strong support for stable operation and optimized scheduling in practical production scenarios.