Comparative Analysis of Deep Learning and ARIMA-based Models for Multi-Step Weather Forecasting in the city of Yaoundé, Cameroon      

Background : The environmental imbalance caused by climate change has significantly complicated weather forecasting, a crucial activity in modern agriculture. Recent research indicates that there is no universally ‘best’ climate prediction model. Interestingly, non-deep learning models have sometimes outperformed deep learning models, which are typically considered highly effective in weather forecasting. Consequently, it is essential to analyze each specific location to determine the most suitable weather forecasting method. Method : In a study focused on the city of Yaoundé, Republic of Cameroon, the performances of various models were evaluated for multi-step weather forecasting. These models included the autoregressive integrated moving average (ARIMA), one-dimensional convolutional neural network (1D-CNN), multi-channel CNN, multi-headed CNN, Vanilla long short-term memory recurrent neural network (LSTM), encoder-decoder LSTM, and encoder-decoder CNN-LSTM. Results : The multi-channel CNN model, with a root mean square error (RMSE) of 3.84°C, outperformed the encoder-decoder CNN-LSTM model, which had an RMSE of 3.90°C, for multi-step univariate weather forecasting. Additionally, the ARIMA (7, 0, 0) model demonstrated strong performance for multi-step multivariate weather forecasting in Yaoundé, with a maximum deviation of 0.79°C, which is less than an RMSE of 1°C. Conclusion : The multi-channel CNN model is recommended for univariate multi-step weather forecasting in Yaoundé, while the ARIMA (7, 0, 0) model is recommended for multivariate multi-step weather forecasting in Yaoundé. The data and scripts used for this study can be accessed at github.com Keywords: ARIMA, CNN, LSTM, multi-step weather forecasting, Yaoundé