Retrieval of missing values in water temperature series using a data-driven model

A measurement buoy with attached sensors has been deployed at our study area to monitor hydrodynamics, water properties, and water quality conditions. High-resolution temporal data have been collected and streamed into an online system that is accessible in nearly real-time. However, in certain circumstances the sensors may fail to provide continuous and high quality data. This results in gaps or corrupted values. The aim of this study was to reconstruct the faulty values. This paper proposes a method based on a data-driven model, using an Artificial Neural Network combined with a Genetic Algorithm to generate a synthetic data series. The generated data can be used as a patch for the incomplete measured data. Additional improvements were achieved by removing seasonal patterns from the original time series using a wavelet decomposition prior to the data-driven model training process. Comparisons with a standard missing-data imputation method using the Kohonen self-organizing map were made to further asses the performance of the proposed data-driven model. The algorithm was applied to water temperature data, but the same approach is applicable to other parameters of interest.