Machine Learning-based Sensor Data Forecasting for Precision Evaluation of Environmental Sensing
Arturs Kempelis, Marta Narigina, Eduards Osadcijs, Antons Patļins, Andrejs Romānovs
Abstract
This paper considers various models for forecasting environmental sensor data values. The aim is to evaluate and compare the performance of forecasting methods, such as machine learning and neural networks when forecasting CO <inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</inf> , Temperature and Humidity sensor data. The research methodology entails finding and employing widely used algorithms to conduct experiments aimed at forecasting humidity, temperature, and CO <inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</inf> sensor data. The models Long Short-Term Memory (LSTM), Convolutional Neural Network (CNN), and Vector Autoregressive (VAR) model were implemented and used in the experiments. The findings reveal that the LSTM model demonstrates the lengthiest training duration but has consistent performance across all evaluation metrics. In contrast, the VAR model excels in temperature forecasting with reduced training times but exhibits inferior performance in forecasting humidity and CO2 levels. The CNN model, however, consistently underperforms in comparison to the other two models, particularly in humidity and CO2 forecasting. Results show that model selection is contingent upon the specific problem and data characteristics, with LSTMs being more for scenarios with long-range dependencies, and VAR models being advantageous for linear and stable relationships between variables.