Through the integration of multimodal data fusion technology and computer AI technology, people’s needs for intelligent life can be better met. This paper introduces the alignment and perception algorithm for multimodal data fusion, which is based on combining the multimodal data fusion model. Taking the air pollutant concentration prediction as an example, the time series of air pollutant concentration is obtained through the LSTM model for concentration prediction, and the attention mechanism is introduced to establish the numerical prediction model of air pollution. Different stations are also selected to acquire weather image data, and the TS-Conv-LSTM multimodal spatio-temporal fusion model of air quality images is constructed by utilizing the Conv-LSTM cell as an encoder, and then the TransConv-LSTM cell, which integrates the anti-convolution and the long-short-term memory network cell, as a decoder. The Gaussian regression model was then used to combine numerical prediction and image prediction models, thus achieving the multimodal synergistic prediction of air pollutant concentrations. The RMSE of the ATT-LSTM model on the dataset was reduced to 8.03 compared to the comparison model, and the predictive fit to the image dataset was above 0.75 for all R² values. The lowest MAE value obtained by the multimodal collaborative prediction model is only 3.815, and the highest R² value is up to 0.985. Introducing deep learning techniques into multimodal data fusion helps to explore the value of massive data more deeply and obtain more comprehensive and reliable information about it.