PURPOSES : Despite the availability of larger traffic data and more advanced data collection methods, the problem of missing data is yet to be solved. Imputing missing data to ensure data quality and reliability of statistics has always been challenging. Missing data are imputed via several existing methods, such as autoregressive integrated moving average, exponential smoothing, and interpolation. However, these methods are complicated and results in significant errors.
METHODS : A deep-learning method was applied in this study to impute traffic volume data of the South Korean national highway. Traffic data were trained using the long short-term memory method, which is a suitable deep-learning method for time series analysis.
RESULTS : Three cases were proposed to estimate the traffic volume. In the first case, which represented the general conditions, the mean absolute percentage error (MAPE) was 12.7%. The second estimation case, which was based on the opposite traffic flow, exhibited a MAPE of 17%~18%. The third case, which was estimated using adjacent-section data, had a MAPE of 18.2%. CONCLUSIONS : Deep learning may be a suitable alternative data imputation method based on the limited site and data. However, its application depends on the specific situation. Furthermore, deep-learning models can be improved using an ensemble method, batch-size, or through model-structure optimization.

ABSTRACT
1. Introductuon
    1.1. Research Background
    1.2. Research Objectives
2. Background
3. Methodology
    3.1. Long Short-Term Memory
    3.2. ATR(Automatic Traffic Recorder) Site andData Condition
    3.3. Deep-Learning Model and Input Features
4. Results
    4.1. Case 1
    4.2. Case 2
    4.3. Case 3
5. Conclusion
REFERENCES

• Jung YooSeok(Korea Institute of Civil Engineering and Building Technology) | 정유석
• Jung Do Young(Korea Institute of Civil Engineering and Building Technology) | 정도영
• Oh JuSam(Korea Institute of Civil Engineering and Building Technology) | 오주삼 Corresponding Author