IAQ-STL-ML: A novel indoor air quality prediction pipeline using meta-learning framework with STL decomposition

Abstract

Today, as people spend over 90% of their time indoors, indoor air quality is crucial due to the health effects of various pollutants. Accurate indoor air pollution predictions can alert occupants to improve indoor air quality before it deteriorates, which can greatly benefit the comfort, health, and safety of indoor occupants. Many recent studies have used deep learning methods to predict air quality. However, these traditional approaches require a large amount of dataset, are difficult to capture the spatial-temporal characteristics of air quality data collected from multiple regions. And selecting the most suitable prediction model based on data characteristics is also an important issue. To address such problems, we propose a novel indoor air quality prediction pipeline (IAQ-STL-ML), which integrates the seasonal trend decomposition using the Loess (STL) and meta-learning framework. In the proposed IAQ-STL-ML pipeline, we first use the STL decomposition method to remove the residual component from the indoor air quality (PM 10 ). Then, meta-features are extracted from the time series data, and based on these meta-features, the meta-learner assigns weights to the predictions of base forecasters and combines these values to predict the PM 10 concentration one hour later. In this study, we solved the “prediction lag” problem by using the STL method on time series PM 10 data. The proposed IAQ-STL-ML pipeline was applied to indoor air quality dataset collected from various regions in South Korea. Experimental results showed that proposed IAQ-STL-ML outperformed the benchmark models with an accuracy of 94.93% and RMSE of 1.876. The overall structure of proposed IAQ-STL-ML pipeline • In this paper, we propose a novel indoor air quality prediction pipeline (IAQ-STL-ML), which integrates the seasonal trend decomposition using the Loess (STL) and deep meta learning framework. • The “prediction lag” problem was solved by removing the residual component from the time series indoor air quality (PM10) using the STL method, and the most correlated input variables were used through correlation analysis. • In this paper, meta-learner was designed to take meta-features as input and automatically assign appropriate weights to several base forecasters, and then combines these values to predict the PM10 concentration one hour later. • The proposed IAQ-STL-ML pipeline was applied to indoor air quality datasets collected from various regions in South Korea, and experimental results showed that the IAQ-STL-ML pipeline outperformed the benchmark model with an accuracy of 94.93% and RMSE of 1.876.

Topics & Concepts