Dynamic forecasting model for indoor pollutant concentration using recurrent neural network

Abstract

Accurate and reliable indoor pollutant concentration prediction is essential to solve the time-lag problem of indoor air quality control systems. Thus, the representation of time in pollutant forecasting models is very important. One approach is to introduce an Elman neural network using a direct inference strategy into the time series forecast of indoor pollutant concentration. In this study, measurements of CO2 (ppm), total volatile organic compounds (mg/m3), particulate matter with a diameter smaller than 2.5 µm (PM2.5; µg/m3), the indoor dry bulb temperature (°C) and relative humidity (%) were carried out in a classroom at a middle school in Beijing, China. To identify air pollution antecedents, input selection was conducted based on correlation analysis. The results show that the information provided by the PM2.5 time series can better simulate the dynamic relationship between input and output data ([Formula: see text]= 0.963 and R2 = 0.928). In addition to the overall goodness of fit ([Formula: see text] = 0.982) of the CO2 time series, the peak and valley prediction capability of the model was evaluated using the relative peak error ( RPE) metric. Information from the valleys of the CO2 time series gives good results ([Formula: see text]). Therefore, a dynamic forecasting model with a direct inference strategy is a capable tool for identifying proper air pollution antecedents.

A Review of Artificial Neural Network Models Applied to Predict Indoor Air Quality in Schools

“…However, the model performance outside the extreme range is not validated. For example, the maximum CO 2 concentration in [114] is 551 ppm, while in other papers, it exceeded 1000 ppm [112,116,117,119].…”

“…However, teaching activities were not included in input variables due to the limitation of data collection. Hu et al [116] set four cases with different combinations of input parameters and found the accuracy was better using only historical concentration data. Notably, they also used sensitivity analysis to calculate the time length of the input variables and found a better accuracy using five timesteps (t, t−1.…”

“…Excluding simulation studies, pollutant concentrations were obtained from real-time sensor monitoring. Most models conducted stepwise prediction of pollutants concentrations [113,114,116,117,119] and used multiple variables as outputs, while Rastogi et al [112] and Elbayoumi et al [115] used daily average PM 2.5 and PM 10 as the output variables.…”

“…Current neural network models for predicting IAQ vary widely regarding algorithm structure and selection of variables. Most of the research has used traditional ANNs [48,[112][113][114][115][117][118][119], while RNNs may be more advantageous for time-series problems [116]. The differences in school location, season, monitoring costs, and the building itself caused inconsistent results in the choice of variables in the nine papers.…”

“…The performance of black-box models is usually indicated by the difference between the predicted and actual values. Hu et al [116] raised the problem of matching computational speed with dynamic control systems; however, their research still focused on the accuracy of the predictive model rather than the environmental benefits of the model. In fact, The models are rarely used in actual buildings due to the constraints of cost and building access, which results in current research being limited by monitoring costs [131] and computational costs [132].…”