Accurate meteorological data, particularly temperature, are crucial across a wide range of fields, from weather forecasting to industrial and research needs.
The Slovenian Environment Agency faces the challenge of missing temperature measurements in their data. This is why our thesis aimed to develop and test various machine learning models that would replace missing values with minimal error. We deal with the event when only temperature data is missing at a given station, while other data remain accessible. We trained and tested our models on 97 stations located across Slovenia. Data from 2016 until 2022 were used for training while data from 2023 and 2024 were used for testing.
Models based on convolutional neural networks showed the most promising results achieving a mean absolute error of 0.4°C for outages lasting up to 2.5 hours.
|
