Predicting CO levels using LSTM and Rolling-Features

Abstract

Accurately predicting carbon monoxide (CO) levels is essential for effective environmental monitoring and safeguarding public health. This research investigates the use of Long Short-Term Memory (LSTM) networks for forecasting CO concentrations specifically evaluating how different learning rates influence model performance. The study aimed to assess the effects of adjusting the learning rate to 0.001, 0.0001, and 0.0005 on the model's accuracy and rate of convergence. A dataset of CO measurements was utilized with feature engineering applied to include lag-based and rolling window features. Results indicated that a learning rate of 0.001 produced the most accurate predictions, achieving the lowest error metrics such as Mean Squared Error (MSE), Root Mean Squared Error (RMSE), Mean Absolute Error (MAE), and Mean Absolute Percentage Error (MAPE). Conversely, smaller learning rates resulted in higher error rates, reflecting slower convergence and less accurate predictions. These findings underscore the importance of selecting the correct learning rate for optimal model performance and suggest that future studies could further investigate learning rate optimization and integrate additional data to improve prediction outcomes.