|
ИСТИНА |
Войти в систему Регистрация |
ИСТИНА ИНХС РАН |
||
Neural Forecasting of the Average Air Temperature Anomaly in the Arctic Region Considering the Cyclonic Activityстатья
- Авторы: Solovyev D.A., Razorenova O.A., Nefedova L.V.
- Сборник: Processes in GeoMedia — Volume VI
- Серия: Springer Geology
- Том: 6
- Глава в коллективной монографии
- Год издания: 2023
- Издательство: Springer International Publishing AG
- Местоположение издательства: Cham, Switzerland
- Первая страница: 401
- Последняя страница: 411
- DOI: 10.1007/978-3-031-16575-7_37
- Аннотация: The article reveals the issue of the neural networks’ applicability, considering the possibility of “mixing” the data set of cyclonic activity indicators in the North Atlantic as parallel inputs for constructing a long-term forecast of the air temperature anomaly in the Arctic region. The importance of considering this factor in the implementation of neural forecasting is determined by the fact that the North Atlantic cyclones provide the supply of heat and moisture to the polar regions and can also modify the ice cover of the Arctic Ocean. A nonlinear autoregression model with external inputs was used as a primary neural network. The training was carried out on data changes in temperature anomalies with a monthly resolution with different observation durations. The training sample is composed of three 20 year vectors, randomly selected from the available set. The software environment STATISTICA Automated Neural Networks was used as a tool in work, which allows researching data preparation to testing a trained neural network. Among all the considered variants of the network parameters, the best results of training the neural network were obtained with 15 hidden layers and a signal delay of 2 samples (simulation clock). Testing the predictive properties of the obtained neural network, carried out on a selection of unused data in training, and showed quite good results. In all cases, the correlation coefficient between the input data and the forecast was more than 0.85. The root-mean-square deviation (model error) did not exceed 0.5 °C.
- Добавил в систему: Нефедова Людмила Вениаминовна