Performance Evaluation of AI-Based Time Series Prediction Models for Predicting Odorous Substance Concentrations in a Residential-Industrial Complex

Odor is an air pollutant that causes discomfort and is increasingly becoming a source of public complaints, particularly around urban industrial complexes. As such, there is a growing need for technologies and management systems that can predict and manage odor occurrences in advance. This study investigates odor prediction using three years of TVOC and ammonia data collected from the Buk-Daejeon area, with the following three objectives: (1) to compare prediction performance across different types of models, including machine learning models (XGBoost, LightGBM) and deep learning models (LSTM, GRU); (2) to analyze the impact of variations in input sequences and prediction targets; and (3) to assess the effect of single-variable versus multi-variable input data on model performance. All models were tuned using Optuna and trained, validated, and tested using an 8:1:1 data split. The analysis showed that the LSTM model consistently exhibited stable predictive performance across most scenarios. Results regarding input sequences and prediction targets revealed that extending the input sequence from 24 hours to 48 hours generally improved prediction accuracy, whereas increasing the prediction horizon from 1 hour to 2 hours tended to degrade model performance. While the multi-variable prediction models outperformed single-variable models at some locations, they did not show a clear overall improvement. These findings highlight the importance of appropriately selecting and structuring input variables when developing odor prediction models. This study provides a foundational analysis and evaluation of odor prediction models and is expected to contribute to the future development of odor management systems and technological advancements.