A Novel Multi-Speed Dilated Convolutional Deep Neural Network Architecture for Meteorological Forecasting

This paper presents a novel Multi-Speed Dilated Convolutional Deep Neural Network (MSDC) architecture for accurate weather forecasting. The MSDC model employs a multi-scale parallel convolutional architecture to capture local features and long-term temporal dependencies in time series data. The model ingests input time series through multiple parallel branches, each applying causal convolutions with varying dilation rates to model the input at different temporal resolutions. The output from the parallel convolutional branches then passes through a multi-head self-attention layer, allowing the model to attend to information from different positions and representational subspaces jointly. In parallel to the convolutional branches, the input time series also passes through a recurrent LSTM layer, and the combined output integrates features gathered by both convolutional and recurrent streams. The MSDC model was evaluated on a dataset comprising historical meteorological data from weather stations in Apulia, Italy, to predict rainfall, air temperature, and wind speed. The model's performance was compared against shallow learning algorithms (SVR and Random Forest) and deep learning models (TCN and Standard Transformer). The results demonstrate that the MSDC model consistently outperforms other models across various weather stations, achieving the lowest MAE and RMSE while maintaining a competitive R2 value. The multi-scale convolutional architecture with parallel branches having different dilation rates and the self-attention mechanism and LSTM provide an apt configuration for learning robust representations from sequence data and effectively capturing local and global patterns for accurate weather forecasting.