Please use this identifier to cite or link to this item:
https://www.um.edu.mt/library/oar/handle/123456789/107897
Title: | Artificial intelligence in short-term meteorological forecasting |
Authors: | Zammit, Ethan (2022) |
Keywords: | Weather forecasting Artificial intelligence Convolutions (Mathematics) |
Issue Date: | 2022 |
Citation: | Zammit, E. (2022). Artificial intelligence in short-term meteorological forecasting (Bachelor's dissertation). |
Abstract: | Air temperature affects each and every one of us, wherever we are located in the world. For this and many other reasons, researchers have long targeted forecasting the weather, aiming for its prediction with the highest possible accuracy. The main challenge in forecasting lies within the chaotic variation of temperature. As a result, most classical techniques are unable to account for the nonlinear relationships between meteorological parameters, and numerical models typically require supercomputer-scale processing power, making them inefficient and inconvenient. In recent years, numerous advancements in Artificial Intelligence have influenced many domains, including meteorology. Through these developments, meteorological researchers managed to achieve substantial improvements upon existing forecasting systems. In this dissertation, an investigation was carried out to explore the limits of the LSTM and TCN AI models. Through an exhaustive search of data and model parameters, we sought to better understand model requirements and capabilities when applied to local air temperature hourly forecasting. Two years of observation data on Marsaxlokk and six years of hourly data on Luqa were used to train and evaluate our models. The LSTM and TCN models were able to outperform three baseline techniques and an additional regression model. Although the TCN provided satisfactory performance, the LSTM was generally found to be superior. The minimum lookback period was found to be 12 hours, yet further addition provides diminishing returns. On the other hand, the horizon was very revealing, identifying two crucial types of performance degradation with forecast horizon size increase. Furthermore, the additional consideration of meteorological parameters as inputs was found to be beneficial, with wind showing the greatest individual benefit. Finally, we recommend that at least 12 months of training data be used, as sharp accuracy deterioration starts occurring in smaller training data sizes. In our opinion, these results extend similar studies, finding that AI techniques can indeed be used to forecast air temperature accurately. In conclusion, we believe that AI-based solutions are the sweet spot between accuracy and efficiency. |
Description: | B.Sc. IT (Hons)(Melit.) |
URI: | https://www.um.edu.mt/library/oar/handle/123456789/107897 |
Appears in Collections: | Dissertations - FacICT - 2022 Dissertations - FacICTAI - 2022 |
Files in This Item:
File | Description | Size | Format | |
---|---|---|---|---|
2208ICTICT390900014431_1.PDF Restricted Access | 2.05 MB | Adobe PDF | View/Open Request a copy |
Items in OAR@UM are protected by copyright, with all rights reserved, unless otherwise indicated.