|
[20]
|
MAE, RMSE, IA
|
Feasibility and practicality were verified experimentally for forecasting PM2.5 using their proposal
|
Algorithmic predictions did not follow real trend accurately and were a bit shifted and disordered
|
|
[22]
|
MAE, RMSE
|
CNN-GRU and CNN-LSTM worked better for PM10 and PM2.5, respectively
|
Hybrid models weakly predicted future highest and lowest levels of PM2.5
|
|
[23]
|
RMSE, R2
|
After using multiple algorithms, it was found that Extra Trees gives the best performance
|
The study was limited in the number of machine learning algorithms compared. There was a bit of a shift between actual and predicted values for most algorithms
|
|
[24]
|
RMSE, MAE
|
CNN could extract air quality features, shortening training time, whereas LSTM could perform prediction using long-term historical input data
|
More evaluation parameters, stating closeness to real values like R2 or IA rather than only errors metrics, could have been used to confirm their models' performance
|
|
[25]
|
NMSE, FB and FA2
|
PM2.5 concentration was predicted using meteorological parameters and PM10 and CO without a history of PM2.5 itself
|
More machine learning models could have been used to test their methodology further
|
|
[26]
|
SMAPE
|
The ensemble of the three models (AccuAir) proved to be better than the individual components tested
|
They did not use LSTM in their Seq2Seq model, although it was proven to be very efficient in time series prediction
|
|
[27]
|
RMSE, MAE and MAPE
|
Their model was compared to Multilayer Perceptron (MLP) and LSTM models and proved to be more stable and accurate
|
Their system predicts only the daily average and cannot be deployed to predict the hourly or real-time concentration of PM2.5
|
|
[28]
|
A comparison of prediction values vs. real value using different sample sets
|
Their proposed system uses many sensors to ensure accuracy and minimize monitoring cost. The system is scalable and suitable for big data analysis
|
The study did not use any clear evaluation metric; instead, they presented a comparison of prediction values vs. actual value using different sample sets
|
|
[29]
|
Calculating AQI and comparing two setups with and without measurements flattened and calibration and accumulation algorithms employed
|
They developed a system that saves bandwidth and energy consumption
|
Further processing by the edge can save even more bandwidth and energy consumption. However, no prediction exists on the edge devices or the cloud side
|
|
[30]
|
There is no evaluation metric of their system, only a proof of concept
|
It tackles security issues of that kind of IoT system. Their IoT solution is scalable, reliable, secure and has HA (high availability)
|
The system is used primarily for monitoring rather than conducting prediction of future pollution levels. It relies on central management and central prediction rather than performing prediction on edge devices
|
|
[31]
|
RMSE, MAE and F1
|
It comprised both prediction and classification to make an alarm system. LSTM was compared to SVR as a baseline, and LSTM was proven to be a better algorithm
|
Their research did not include a comparison to other works and used only one base model
|