LSTM for NOx Time Series Prediction

Subproject for Stockholm Environmental Institute (AQsensor: air quality and pollutants prediction of dynamic traffic using emerging IoT sensors)

Supervisor: Xiaoliang MA, KTH Royal Institute of Technology

Task

Develop and train LSTM model to predict the hourly NOx concentration data in Stockholm for the next 3 days and evaluate the accuracy compared with other models.

Methodology & Procedure

Data Inputs

• Inputs of Univariate Model

The inputs only contains NOx concentration.

• Inputs of Multivariate model

Except the NOx data, the inputs also contains environmental data such as Difftemp, Global radiation, STD WD, STD WS, STD VertWind, Temp, WD and WS.

Data Analysis & Processing

• Time lag acf & pacf, FFT
• Outliers detection: Density curve, Box-plot
• Drop nan
• Data normalization

Models Construction

Univariate Model

Model	Notes
LSTM	keras
Prophet	fbprophet
Xgboost	xgboost
ARIMA	statsmodels
SVM	sklearn
KNN	sklearn
Bayes	sklearn
DecisionTree	sklearn

Multivariate model

Features	Notes
LSTM	keras
DecisionTree	sklearn
Xgboost	xgboost

Training

Data of 1.1-12.31/2015 as train, 1.1-1.3/2016 as validation

Evaluation

The prediction and the ground truth of next 3 days are visualized in a figure, and RMSE, MAE, MAPE, MedAE, r2_score and explained_variance_score are used for evaluation