PM2.5 prediction based on attention mechanism and Bi-LSTM

Abstract

Abstract This study enhances the Bi-LSTM model by incorporating an attention mechanism, which could provide the model with stronger data generalization capabilities. Moreover, it can predict a broader range of data and exhibits enhanced handling and adaptability to anomalies. Through the utilization of the attention mechanism, this research partitions the weights of the feature values, precisely dividing the input LSTM's feature values based on their weights. This enables the Bi-LSTM to more accurately capture relationships between different feature values in time series and dependencies on various features. Given the diverse air quality conditions in different regions, the introduced attention mechanism in Bi-LSTM manages the weights of different feature values. The Bi-LSTM, enhanced with attention mechanisms, excels at handling relationships in time series data, allowing it to predict PM2.5 values in more complex air quality environments. It demonstrates improved capabilities in handling anomalies. Even in air quality scenarios with various complex conditions, the model maintains satisfactory predictive quality.