OzoneNet:A spatiotemporal information attention encoder model for ozone concentrations prediction with multi-source data

Surface ozone (\(O_3\)) pollution is a serious environmental problem that endangers human health, and it is also an increasingly prominent environmental problem in the World. Existing works focus on how to directly improve the accuracy of predicting the target sequence from the input sequence while ignoring the inherent uncertainty of ozone in the atmosphere during the modeling process. Therefore, we utilize data fusion techniques to integrate ground observation data, satellite data, and reanalysis data for simulating atmospheric dynamics and enhancing prediction accuracy. We developed a sequence to sequence using a unit embedded with spatiotemporal information self attention mechanism as its encoder (OzoneNet) predict ozone concentration in the future. In the proposed method, we utilize the LSTM model with Spatiotemporal information self-attention mechanism to extract fixed Spatiotemporal data features, and the temporal dimension characteristics in long-term series are modeled by sequence-to-sequence network. Results show that the model has higher reliability and validity, outperforming benchmark models in simulating future changes in \(O_3\) concentrations. The progeress of this method can help the public take corresponding protective measures, provide scientific guidance for the government’s coordinated control of regional pollution, and can also provide important references for environmental protection and climate change research