Efficient and Reproducible Temperature Forecasting Using CNN-BiGRU Based Architectures

This research presents a reproducible and structured approach for predicting daily maximum temperatures using deep learning models. The focus is on ensuring fair model comparison and reliable predictions. Weather data is collected and preprocessed to fix missing values, incorrect entries, and timing errors, which helps maintain the quality of the time-series data. The cleaned data is then framed as a supervised learning problem using a sliding window method, where the past 10 days of temperature values are used to predict the next day’s temperature. This approach captures the time-based relationships in the data while keeping the models efficient. The dataset is split into training, validation, and testing sets in chronological order to avoid data leakage and to properly assess model performance. Eight different deep learning architectures Vanilla RNN, LSTM, GRU, CNN, GRU-LSTM, BiGRU-RNN, BiGRU-LSTM, and CNN-BiGRU are implemented using Keras with TensorFlow as the backend. All models use the same hyperparameters to allow fair evaluation. The proposed method provides a robust, transparent, and reproducible framework for temperature forecasting using time-series data.