University of Cambridge > > Rainbow Group Seminars > Bayesian Deep Learning to Predict Air Pollution & Personalized Air Pollution Monitoring and Health Management

Bayesian Deep Learning to Predict Air Pollution & Personalized Air Pollution Monitoring and Health Management

Add to your list(s) Download to your calendar using vCal

If you have a question about this talk, please contact Hatice Gunes.

HKU and Cambridge colleagues, working together in HKU -Cambridge CEERP , have recently won a major grand challenge grant on AI and Air Pollution Monitoring and Health Management. These two talks will be providing insight into the research work conducted so far as part of this project.

The project addresses five major challenges. FIRST , urban air quality data is sparse, rendering it difficult to provide timely personalized alert and advice. SECOND , collected data, especially those involving human inputs, such as health perception, are often missing and erroneous. THIRD , data collected are heterogeneous, and highly complex, not easily comprehensible to facilitate individual or collective decision-making. FOURTH , the causal relationships between personal air pollutants exposure (specifically PM(2.5,1.0) and NO2 ) and personal health conditions, and health (well-being) perception, of young asthmatics and young healthy citizens, are yet to be established. FIFTH , one must determine if information and advice provided can effect behavioral change.

Air pollution has deteriorated rapidly in many metropolitan cities, such as Beijing. Since poor air quality has clear public health impacts, accurately monitoring and predicting the concentration of PM2 .5 and other pollutants have become increasingly crucial. This talk presents a hybrid approach where time series decomposition and Bayesian Long Short-Term Memory (BLSTM) are combined as a framework for air pollution forecast, based on historical data of air quality, meteorology and traffic in Beijing. LSTM has been proven to achieve state-of-the-art performance in many time series prediction applications due to its capability of memorizing long term sequential correlations. In addition, the model uncertainty estimates generated by Bayesian methods may reduce overfitting, improving the accuracy of the prediction. In our experiment, deseasonalized features are fed into BLSTM to predict the air pollution in the next 48 hours of each monitoring station in Beijing. Results show that the BLSTM framework outperforms the baseline models including SVR , STL, ARIMA , and traditional LSTM with dropout regularization.

This talk is part of the Rainbow Group Seminars series.

Tell a friend about this talk:

This talk is included in these lists:

Note that ex-directory lists are not shown.


© 2006-2018, University of Cambridge. Contact Us | Help and Documentation | Privacy and Publicity