For multi-step forecasting of traffic flow, a convolutional neural networks (CNN)-gate recurrent unit (GRU) model based on encoder-decoder (ED) framework was proposed, referred to as the ED CNN-GRU model. In this model, CNN serves as the encoder, capturing information from the traffic flow sequence, which is then interpreted and outputted by the GRU decoder. Experimental results show that compared with CNN and GRU models, ED framework effectively solves the problem of rapid error accumulation. Compared with other benchmark models, CNN and GRU models are superior in feature extraction and interpretation of traffic flow series. In terms of the traffic flow prediction task of 12 steps in the future, compared with other benchmark models, the root mean square error of the univariate input ED CNN-GRU model is reduced by about 0.344 to 6.464, and the mean absolute error is reduced by about 0.192 to 0.425. Additionally, compared with univariate input, the ED CNN-GRU model with multivariate input exhibits a better fitting performance. These findings confirm that ED CNN-GRU model possesses strong forecasting capabilities for multi-step traffic flow forecasting tasks with varying input dimensions, and provides a multi-step traffic flow forecasting model that supports both univariate and multivariate input for cities with diverse data acquisition conditions.