PM2.5 Concentration Prediction Based on CNN-BiLSTM and Attention Mechanism

The concentration of PM2.5 is an important index to measure the degree air pollution. When it exceeds standard value, considered cause pollution and lower quality, which harmful human health can a variety diseases, i.e., asthma, chronic bronchitis, etc. Therefore, prediction helpful reduce its harm. In this paper, hybrid model called CNN-BiLSTM-Attention proposed predict over next two days. First, we select data in hours from January 2013 February 2017 Shunyi District, Beijing. auxiliary includes quality meteorological data. We use sliding window method for preprocessing dividing corresponding into training set, validation test set. Second, composed convolutional neural network, bidirectional long short-term memory attention mechanism. parameters network structure are determined by minimum error process, including size convolution kernel, activation function, batch size, dropout rate, learning determine feature input output evaluating performance model, finding out best 48 h. Third, experimental part, set check on prediction, compared other comparison models, lasso regression, ridge XGBOOST, SVR, CNN-LSTM, CNN-BiLSTM. conduct (48 h) long-term (72 h, 96 120 144 h), respectively. results demonstrate that even predictions h with better than models terms mean absolute (MAE), root square (RMSE), coefficient determination (R2).