This paper describes our approach for the MediaEval2021 “CrossData Analytics for (transboundary) Haze Prediction” subtask1. The objective of this subtask is to predict PM10 values at diferent locations in multiple countries using data only from each country itself. In addition, we have applied XGBoost to deal with missing PM10 values on the training dataset and Long Short-term Memory (LSTM) [2] models to predict air pollution.
Nowadays, air pollution leads to increasing cases of cardiovascular and respiratory diseases. It also afects social and economic activities. By using data from the last several days to predict air pollution for upcoming days, we can plan appropriate activities to protect our health.
As given in the task description [
Each data point in the training dataset represents the information
of a location of one day in the country. We observed many missing
values in the datasets; thus, we decided to employ two diferent
methods to impute the missing values according to the value type.
For data extracted from weather stations (i.e., temperature, rainfall,
humidity, and wind speed), we filled the missing values with the
mean values of its stations. For data related to air quality from
monitoring stations, i.e., PM10, we employed XGBoost [
First, the missing values of weather features on the training dataset were filled using the first method. Next, we created a new dataset from the original training dataset by dropping the rows where PM10 values are missing. Then, the new dataset was used to build the XGBoost model to predict missing PM10 values on the original training data from weather features.
It is worth noticing that all-weather features of Thailand collected in 2015 are missing. Therefore, to avoid interference when iflling missing values, we dropped all data points in that year. 2.2
Research studies have shown that LSTM is suitable for time series
data [
2.2.1 LSTM. One of the disadvantages of Recurrent Neural
Network (RNN) is it can not process long sequences; LSTM architecture
is proposed to solve that problem. An LSTM cell, as shown in
Figure 1, has a cell state C that allows the information to flow through
for long-term memory. It also includes three gates: forget
gatedecide what information should be kept or discarded by looking
at the previous state and current input. Input gate decides what
information is essential at the current step and how to add to the
cell state; output gate decides what the output should be.
2.2.2 Bi-LSTM. Bi-LSTM model, which was developed from
Bidirectional Recurrent Network [
2.2.3 Stacked LSTM. Stacked LSTM is a model that includes
multiple LSTM layers. By making the model deeper, it has proved
its efection in sequence data [
For each country, we did not develop multiple models for each
station but combined data from all the stations into one complete
dataset and build models to recognize the pattern form the dataset.
The first 80% of the data is used for training sets, and the last 20%
of the data is used for validation sets. Information of ten previous
days, which includes weather features and air quality, is used to
predict the PM10 values of the upcoming day. The Adam optimizer
[
For Brunei, we experimented with LSTM, Bidirectional LSTM (Bi-LSTM), and Stacked LSTM. The organizers provided the PM10 values hourly for Singapore and Thailand, and we only tested all three LSTM models for three first hours PM10 values (i.e., 10_1, 10_2, and 10_3). The results show that Bi-LSTM is slightly better than LSTM and Stacked LSTM, so we chose to employ the Bi-LSTM model for each hourly PM10 value for the final predictions. 3
We evaluate the proposed models using RMSE metric calculated as follow: √︄ Í
=1 || () − ˆ () ||2 = where is the number of data points, () is the i-th measurement and ˆ () is its ground truth.
The experimental results on validation sets of Brunei, Singapore, and Thailand datasets are shown in Table 1. In Brunei, Bi-LSTM achieves a slightly better result than LSTM, with a score of 3.625. For Singapore and Thailand, the average scores are 5.821 and 10.624.
In this work, Bi-LSTM initially has shown some promising results for Brunei and Singapore on the validation sets. However, it did not perform as expected in the test sets. It might be because our missing values imputation technique is not good enough and the models can’t fully recognize the pattern of the datasets.
Finally, we would like to send our thanks to AISIA Research Lab for supporting our team; the Organization Board of MediaEval 2021 and the Task Organizer for providing us with an opportunity to participate in the competition. Insight for Wellbeing: Cross-Data Analytics for (transboundary) Haze Prediction