The continuous deterioration of air quality, the frequent occurrence and its following adverse effects of air pollution incidents have caused continuous public concerns. Therefore, achieving accurate prediction of air pollutants has significant role for governance efforts. In this paper, a combined weight forecasting model (CWFM) for NO₂ concentration in Beijing is constructed based on three single prediction models. Firstly, the input data are decomposed by discrete wavelet transform to increase the dimensionality of the data. Secondly, DWT-LSTM, DWT-GRU and DWT-Bi-LSTM models are constructed using the wavelet decomposition results and Long short-term memory neural network (LSTM), Gated recurrent units (GRU) and Bi-directional long-short term memory neural network (Bi-LSTM), respectively, followed by comparison and contrast of the prediction results. Finally, the three single prediction models are incorporated into the combined weighted forecasting model by weight assignment. The results show that the combined weight forecasting model constructed in this paper can improve the prediction accuracy by combining the advantages of the every single prediction models, and by contrasting with any single prediction model, the combined model structure is more suitable for the prediction of NO₂ concentration in Beijing.