| Abstract [eng] |
Although the power of supercomputers is constantly increasing, the traditional analytical approaches to weather forecasting are still not considered to be accurate by the general public. In addition to this, the duration of a single forecast computation sometimes can take up to 2 hours. This means that weather prediction models sometimes fail to react to rapid and dangerous changes in weather conditions, which may result in the destruction of material possessions or cost human lives. This master's thesis describes the creation and analysis of precipitation forecast algorithms, which are based on machine learning and artificial neural networks. These algorithms are designed for fast short-term precipitation forecasting over the territory of Lithuania and rely on Dopler’s weather radar images. To create accurate machine learning models and gain more experience in result analysis, related work, which concerns meteorology, precipitation movement, weather radar technology, machine learning, and its applications to weather forecasting is presented. All of the created algorithms have surpassed the accuracy of a simple persistency model, and the convolutional neural network with 4 layers showed the best results. A positive impact of the additional static layers (random and elevation data) in the model input was also observed in the convolutional neural network with 3 layers. Obtained results show the importance of a search for new algorithms in areas that already have stable traditional mathematical methods: if the technological progress in machine learning algorithms continues, weather forecasts will be generated faster and will use less amount of energy. This master's thesis highlights the importance of interdisciplinary dialog between artificial intelligence and meteorology scientist communities. |
