Graph Neural Networks for Missing Data Imputation in Time Series from Meteorological Sensors

Application of Artificial Neural Networks methods for the analysis of sensors data has become a priority nowadays. In this paper, Graph Neural Networks are exploited for the imputation of missing data in meteorological time series. To construct the graphs, a Horizontal Visibility Graphs approach is used. The dataset is populated with the acquisitions of five sensor nodes, deployed for several months in distinct vineyards covering a 500 m × 300 m area near the city of Siena, Italy. The monitored physical quantities are air temperature and relative humidity, leaf wetness, solar radiation and amount of rain precipitation; the sampling is performed approximately every five minutes. The obtained results showed good performance for the prediction of missing values considering several months of tests and different nodes, obtaining low Root Mean Squared Error and Mean Absolute Error.