A0342
Title: A vector error correction model to address sensor-based time series
Authors: Maria del Carmen Robustillo Carmona - Universidad de Extremadura (Spain) [presenting]
Lizbeth Naranjo Albarran - Universidad Nacional Autonoma de Mexico UNAM (Mexico)
M Isabel Parra Arevalo - Universidad de Extremadura (Spain)
Carlos Javier Perez Sanchez - University of Extremadura (Spain)
Abstract: Vector Error Correction (VEC) models are useful for analyzing complex and dynamic long-term relationships between variables under a cointegration approach. These models have been widely used in some areas, such as econometrics, but in a very limited way to time series of sensor data, where they could be very useful. Two particularly interesting cases are i) sensor data from variables related to the status of the beehives in the context of precision beekeeping; ii) sensor data from acoustic features extracted from voice recordings in the context of computer-aided diagnosis systems for detecting Parkinsons' disease. Several experiments have been conducted to assess the performance of this model, and comparisons with linear multivariate state-space models have been made. A precision beekeeping dataset obtained from the we4bee database is shown as an example. Four inner temperatures, humidity, and weight, regularly collected from four beehives, led to four multivariate time series. Mean absolute errors of these variable predictions were estimated for 1, 3, and 7 days in a rolling cross-validation framework. Overall, the VEC model provided better results between 83.3\% and 92.6\% of the predictions made, depending on the beehive. The achieved improvement percentages suggest that considering cointegration in sensor data contexts may provide more competitive models.
