Cloud computing systems require accurate proactive resource allocation to achieve high quality cloud services and efficient cloud resource utilization, and proper proactive resource allocation requires accurate workload prediction. Many existing workload prediction methods applicable to cloud computing systems either have limited prediction accuracy or lack practicality due to excessive overhead. A virtual machine workload prediction algorithm based on dilated convolutional neural network (VMWPD) was proposed to address the above problems. The algorithm adopted a prediction model based on dilated convolutional neural network, applied the mechanism that the prediction model predicts the amount of change in the workload instead of directly predicting the workload, and possessed the ability of real-time prediction and real-time training, which striked a better balance between prediction accuracy and overhead. Evaluation experimental results show that VMWPD improves the accuracy of workload prediction algorithm by 32.45% and reduces the time overhead by 45.10% compared to long short-term memory (LSTM) based workload prediction algorithm. As can be seen, the proposed method significantly reduces the cost while maintaining a certain level of accuracy.