Forecasting climate data for tokat region with deep learning methods

In this thesis, deep learning models were developed to forecast temperature at different time scales—daily, weekly, and monthly—using 25 years of comprehensive meteorological data from Tokat, Turkey. To capture the complex temporal patterns in the data and generate future predictions, advanced recurrent neural network architectures such as Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), and Bidirectional LSTM (BiLSTM) were employed. The modeling process involved meticulous steps including data preprocessing, missing value imputation, normalization, and feature selection. Missing data were filled using linear interpolation and moving average techniques, and input variables were selected based on their completeness and statistical consistency. Each model was trained with different architectural configurations, input feature sets, and time-step scenarios. Their performances were evaluated using key statistical metrics such as Root Mean Square Error (RMSE), Mean Absolute Error (MAE), and the Coefficient of Determination (R²). Results revealed that multivariate models, which incorporate multiple meteorological features as inputs, achieved significantly lower error rates compared to univariate models that rely solely on temperature. Furthermore, forecasts at weekly and especially monthly time scales produced more stable and generalizable outputs than those at the daily scale. Finally, a multi-output BiLSTM model capable of simultaneously predicting all meteorological variables was developed to assess the potential of multi-target forecasting. This study systematically explores the effectiveness of different deep learning architectures in meteorological time series forecasting and demonstrates that long-term predictions can be made using local climate data in Turkey. The findings have the potential to inform future research in climate modeling, early warning systems, and sustainable environmental planning.