KTRojas0212/cnn-bilstm-temperature-togo

Benchmarking 13 models for daily temperature forecasting using ANAMET ground-station data from Lomé, Togo # Explainable Comparative Deep Learning for Daily Temperature Forecasting: Lomé, Togo MSc thesis project in Mathematics (Data Science), Pan African University / Jomo Kenyatta University of Agriculture and Technology (JKUAT), Nairobi. A benchmark of thirteen forecasting models for daily air temperature at Lomé, Togo, spanning naïve, classical statistical, classical machine learning and deep learning families including a hybrid CNN-BiLSTM-Attention architecture. *Status:* in progress. ## Problem Togo has a sparse observational network and limited local computing capacity for numerical weather prediction, yet reliable daily temperature information matters for agriculture, public health and energy planning. Deep learning is often proposed as a low-cost alternative, but published studies rarely test hybrid architectures against strong simple baselines on the same data. This project builds that comparison for a data-scarce West African setting. ## Data - *Source:* Agence Nationale de la Météorologie (ANAMET / Météo Togo) - *Station:* Lomé Airport synoptic station - *Period:* 2015–2026, daily resolution - *Target variable:* daily air temperature (T2M) Ground-station observations, not satellite or reanalysis estimates. Data provenance was documented directly with ANAMET's Climatology Division. The raw records are not redistributed in this repository; access requests go to ANAMET. ## Models compared Thirteen models across four families, all trained and evaluated on identical splits: | Family | Models | |--------|--------| | Naïve | Persistence, Seasonal Naïve | | Classical statistical |Linear Regression, ARIMA(5,0,0) | | Classical machine learning |XGBoost, Random Forest | | Deep learning | ANN, LSTM, BiLSTM, GRU, CNN-LSTM, CNN-BiLSTM, CNN-BiLSTM-Attention | The hybrid *CNN-BiLSTM-Attention* model combines 1D convolutional layers for local feature extraction, bidirectional LSTM layers for temporal dependencies, and an attention mechanism that weights time steps …