<p>Chemical oxygen demand (COD) is a crucial indicator of organic pollution in water bodies. To increase the accuracy of COD analysis and prediction, a model framework was proposed in the paper. A frequency division method, variational mode decomposition (VMD), was used to complete the time domain decomposition of the COD original data before model prediction. The original data were separated into five intrinsic mode functions with different frequency bands, IMF1 to IMF5, based which the influences of meteorological factor and water quality factor on the COD were explored. The long-term changes of COD were influenced largely by nutrient factors, such as phosphorus and nitrogen, while the short-term fluctuations exhibited relatively stable. Then, three models, random forest (RF), long short-term memory (LSTM) and gated recurrent units (GRU) were used to predict the COD with the original data or the signal data processed by VMD as input parameters. The RMSE (root mean square error), MAE (mean absolute error), and SMAPE (symmetric mean absolute percentage error) were taken to evaluate the model prediction effect. The results showed that the prediction performance is as follows: RF &gt; GRU &gt; LSTM and VMD-GRU &gt; VMD-LSTM &gt; VMD-RF. The prediction performance of LSTM and GRU can be significantly improved by frequency division. The improvement rate of the LSTM is 73.45%, 73.01% and 69.58%, respectively, and that of the GRU is 72.11%, 75.19% and 73.14%. Therefore, the VMD-LSTM and VMD-GRU models can be applied to the COD analysis and prediction in the Chengdu area.</p>