Predicting Indoor PM2.5 Concentration using LSTM-BNN in Edge Device

Abstract

Many researchers already perform PM2.5 forecasting. However, the majority of research focuses on predicting PM2.5 concentrations in outdoor environments. In contrast, PM2.5 indoor prediction is rarely conducted, despite being more difficult. This study proposes an LSTM-BNN indoor PM2.5 concentration prediction model. The LSTM in the LSTM-BNN model extracts nonlinear correlations from multivariate time series input while the BNN predicts the PM2.5 concentration. Using multivariable input data, the proposed model estimates PM2.5 values 1 hour, 2 hours, and 3 hours in advance. In addition, the proposed model is compared to RNN, LSTM, Single Dense, Multi Dense, and ConvLSTM. MSE, RMSE, MAE, MAPE, and R2 are employed to evaluate the LSTM-BNN model objectively. The LSTM-BNN model beats other models with 1-hour, 2-hour, and 3-hour prediction MAPE and R2 values of 0.001 and 0.999, 0.004 and 0.996, and 0.004 and 0.999, respectively.