Attention-Based CNN-BiGRU Models for Temporal Temperature Anomaly Detection Using in Climate Time Series

Detecting climate anomalies is crucial for long-term climate change studies and for identifying potential extreme events or irregularities in climate behavior that may reflect anomalies. In contrast to short-term weather forecasting, climate anomaly detection identifies anomalous departures from historical climate patterns at a longer timescale. This research introduces an innovative Attention-based CNN-BiGRU Hybrid approach for detecting temporal temperature anomalies in climate time-series. It combines CNN's ability to extract local temporal features with BiGRU's ability to recognize patterns in a time series over long intervals; and incorporates an attention layer so that it focuses on those time points that are most important for identifying anomalous temperatures; thereby enhancing its performance and interpretation. The model was trained on surface temperature indicators collected from multiple countries during the period from 1961 to 2024. A number of pre-processing and feature engineering techniques were applied to enhance temporal sensitivity to anomalies, including missing value imputation, normalization, lag features, and year-over-year change computation. The proposed approach was assessed using R², RMSE, MAE and Pearson correlation. The experimental results showed that the Attention-based CNN-BiGRU model performed better than the baseline LSTM and GRU models for detecting significant trend deviations in temperature. The presented framework provides a scalable and domain independent solution for climate time-series anomaly detection, and can be extended to other sequential data applications.