Arctic and Antarctic Research Institute
Saint Petersburg, St. Petersburg, Russian Federation
In the face of ongoing climate change, Arctic regions are becoming increasingly vulnerable to extreme hydrological events such as floods. Traditional forecasting methods are often unable to capture complex nonlinear and long-term dependencies in hydrometeorological data, posing significant risks to the management of coastal technosphere safety. The objective of this study was to develop and test a model for forecasting daily water levels in the Arctic Pur River at the Samburg outlet during the navigation season, based on a transformer neural network architecture. This model aims to improve the accuracy and lead time of flood warnings. The study utilized a unique Transformer-type neural network architecture adapted for time series forecasting. The model incorporates a multi-headed self-attention mechanism, positional coding, and fully connected layers. Training and validation were conducted on a multi-year daily data set, including target water level indicators and 12 associated hydrometeorological features. The MAE, RMSE, and MAPE metrics were used to assess accuracy. The constructed model demonstrated high accuracy in forecasting diurnal water level dynamics. The calculated quality metrics (MAE = 10,76, RMSE = 14,00, MAPE = 2,69 %) confirm its effectiveness. The model successfully identifies both seasonal trends and abnormal flood peaks, which is important for early warning systems. The transformer model is a promising tool for solving problems of operational river level forecasting in challenging Arctic conditions. Implementation of the developed approach in risk management practices allows for a transition from reactive to proactive strategies, ensuring timely management decisions and mitigating man-made and natural risks.
water level regime, forecasting, the Pur River, transformer neural networks, navigation safety
1. Harlamp'eva N.K., Tret'yakov M.V., Ivanov V.V. Arkticheskaya ust'evaya gidrologiya: istoriya issledovaniya // Geografiya i kraevedenie v Yakutii i regionah Rossii: materialy Vserossijskoj nauch.-prakt.konf.s mezhdunar. uchastiem, posvyashchennoj 85-letiyu so dnya rozhdeniya G.N. Maksimova. Yakutsk: Izdatel'skij dom SVFU, 2024. S. 115–120. EDN GUMXFY.
2. Resursy poverhnostnyh vod SSSR: Gidrologicheskaya izuchennost'. L.: Gidrometeoizdat, 1964. T. 15. Vyp. 3.
3. Time Series Analysis: Forecasting and Control / G.E.P. Box [et al.] // John Wiley & Sons, 2015. 720 p.
4. Hochreiter S., Schmidhuber J. Long short-term memory // Neural computation. 1997. T. 9. № 8. S. 1735–1780.
5. Learning phrase representations using RNN encoder-decoder for statistical machine translation / K. Cho [et al.] // arXiv preprint arXiv:1406.1078. 2014.
6. Rainfall–runoff modelling using long short-term memory (LSTM) networks / F. Kratzert [et al.] // Hydrology and Earth System Sciences. 2018. T. 22. № 11. S. 6005–6022.
7. Attention is all you need / A. Vaswani [et al.] // Advances in neural information processing systems. 2017. T. 30. № 1. S. 5998–6008.
8. Deep transformer models for time series forecasting: The influenza prevalence case / N. Wu [et al.] // arXiv preprint arXiv:2001.08317. 2020.
9. Informer: Beyond efficient transformer for long sequence time-series forecasting / H. Zhou [et al.] // Proceedings of the AAAI conference on artificial intelligence. 2021. T. 35. № 12. S. 11106–11115.
10. Volkova N.A. Analiz mnogoletnih kolebanij urovnya vody reki Pur dlya obespecheniya bezopasnosti sudohodstva // Prirodnye i tekhnogennye riski (fiziko-matematicheskie i prikladnye aspekty). 2025. № 2 (54). S. 6–22. DOIhttps://doi.org/10.61260/2307-7476-2025-2-6-22
11. Volkova N.A. Kompleksnyj podhod k snizheniyu avarijnosti na vnutrennih vodnyh putyah arkticheskogo regiona Rossii // Izvestiya Peterburgskogo universiteta putej soobshcheniya. 2025. T. 22. № 3. S. 761–775. DOI:https://doi.org/10.20295/1815-588X-2025-3-761-775
12. Volkova N.A., Romashova K.V. Metodika dolgosrochnogo prognozirovaniya maksimal'nyh urovnej vody reki Taz // Tekhnosfernaya bezopasnost' v Arktike: materialy v ramkah VIII Mezhdunar. nauch.-prakt. konf. SPb.: S.-Peterb. un-t GPS MCHS Rossii, 2025. S. 223–226. EDN RFVHDX.
13. Deep quantum-transformer networks for multimodal beam prediction in isac systems / S. Tariq [et al.] // IEEE Internet of Things Journal. 2024. T. 11. № 18. S. 29387–29401.
14. Diformer: A difference transformer network for remote sensing change detection / H. Lin [et al.] // IEEE Geoscience and Remote Sensing Letters. 2024. T. 21. S. 1–5.
15. Nikitin S.D., Mamontov A.I. O szhatii nejrosetej s arhitekturoj «Transformer» // Izmerenie, kontrol', informatizaciya: materialy XHVI Mezhdunar. nauch.-tekhn. konf. / pod red. L.I. Suchkovoj. Barnaul: Izd-vo AltGTU, 2025. 216 c.
16. Klyukovkin G.K. Nejronnye seti protiv klassicheskih ML-modelej: v kakih sluchayah stoit uslozhnyat' arhitekturu // Aktual'nye issledovaniya. 2024. № 37 (219).
17. Attention is all you need / A. Vaswani [et al.] // Advances in Neural Information Processing Systems. 2017. № 30. P. 5998–6008.
18. How to optimize the systematic review process using AI tools / N. Fabiano [et al.] // JCPP advances. 2024. T. 4. № 2. S. e12234.
19. Gidrologicheskij ezhegodnik. L., Gidrometeoizdat, 1936–2012. T. 6. Vyp. 0-9.
20. Svidetel'stvo o gosudarstvennoj registracii bazy dannyh № 2020621470 Ros. Federaciya. «Nauchno-prikladnoj spravochnik «Klimat Rossii»: № 2020620899: zayavl. 09.06.2020; opubl. 18.08.2020 / V.N. Razuvaev, O.N. Bulygina, N.N. Korshunova [i dr.]; zayavitel' federal'noe gosudarstvennoe byudzhetnoe uchrezhdenie «Vserossijskij nauchno-issledovatel'skij institut gidrometeorologicheskoj informacii – Mirovoj centr dannyh». EDN MXYBEY.
