The dynamic weighing signal of dairy cows contains many signals in different frequency domains, including the weight signal of dairy cows, the inertial component signal and various noise signals. In the previous studies, the information utilization rate of dynamic weighing signal was low, and the deep information of weighing signal could not be fully extracted. To solve this problem, a method based on variational mode decomposition（VMD）and long short-term memory network (LSTM) dynamic weighing algorithm was proposed to improve the accuracy of weight prediction. Firstly, the threshold filtering method was used to obtain the effective signal from the collected dairy cow dynamic weighing signal. Secondly, in order to extract the deep information contained in the dynamic weighing signal of dairy cows, the VMD algorithm was used to decompose the pre-processed effective signal into five intrinsic mode functions (IMF). Finally, each IMF component was combined with the effective signal as feature, which was input into the LSTM neural network as features for training, and then the weight of cows was output. The prediction results of models with different characteristics were compared, as a result, the model with the minimum error was selected as the cow body weight prediction model. The experimental results showed that the proposed dynamic weighing algorithm can effectively extract the deep information contained in the dynamic weighing signal of dairy cows. The average relative error of weight prediction was 0.81%, and the root mean square error was 6.21kg. Compared with EMD algorithm and GRU algorithm commonly used in the field of dynamic weighing, the error of the proposed algorithm was smaller.