HYBRID TIME SERIES AND DEEP LEARNING MODELS FOR ACID RAIN PREDICTION
MEETCON – EUROPE 4. ULUSLARARASI YENİLİKÇİ ÇALIŞMALAR KONGRESİ 26-30 EKİM 2026, Madrid, İspanya, 26 - 30 Ekim 2026, cilt.1, sa.1, ss.106-111, (Tam Metin Bildiri)
- Yayın Türü: Bildiri / Tam Metin Bildiri
- Cilt numarası: 1
- Basıldığı Şehir: Madrid
- Basıldığı Ülke: İspanya
- Sayfa Sayıları: ss.106-111
- Eskişehir Osmangazi Üniversitesi Adresli: Evet
Özet
This study aims to develop hybrid models for pH level prediction using monthly acid rain data
from the Amasra station covering the period 2000–2026. A two-stage modeling approach was
adopted. In the first stage, the linear components of the time series were modeled using
ARIMA and SARIMA models. In the second stage, the non-linear patterns that these models
failed to capture were predicted through error terms using LSTM, Support Vector Machines
(SVM), Ridge Regression, and Artificial Neural Networks (ANN). The ARIMA model was
constructed with autoregressive and moving average components, while SARIMA
additionally incorporated a twelve-month seasonal structure. The LSTM model learned
temporal dependencies through its long short-term memory architecture, whereas SVM
captured non-linear relationships using a radial basis kernel function. Ridge Regression
controlled overfitting via a regularization parameter, and the multi-layer ANN employed non
linear activation functions in its hidden layers. The final prediction was obtained by
combining the linear and non-linear components. The findings revealed that ARIMA and
SARIMA models alone were insufficient in explaining pH variability and particularly failed
to predict extreme acid rain events. In contrast, hybrid models demonstrated significant
performance improvements over the base models. The SARIMA-LSTM hybrid model
achieved the highest prediction accuracy, explaining a substantial portion of pH variability
and considerably reducing the average prediction error. The improvement gained through
error term modeling highlights the complementary role of deep learning algorithms when
linear methods alone prove inadequate. Forward predictions generated for the 2026–2027
period indicate a high risk of acid rain, particularly during spring and summer months.
Pollutant source analysis revealed that sulfate is the dominant pollutant in summer months,
while nitrate prevails in winter months. The developed hybrid model can serve as an effective
early warning system for the region