Comparison of Deep Learning Models (CNN and DNN) for Multivariate Time Series Dataset

This study compares the performance of convolutional neural networks (CNNs) and deep neural networks (DNNs) for multivariate time series data. We analyse not only final loss values but also initial learning behaviour and prediction capabilities across different horizons. We trained and evaluated both models on real-world data incorporating consumption, price, day of the week, and month. We analysed loss curves, mean absolute error (MAE), and two-week ahead predictions. Our study goes beyond final accuracy, examining initial learning dynamics and the impact of specific features on prediction horizons. This nuanced analysis provides valuable insights for selecting the optimal model for different practical applications. Both CNNs and DNNs achieved similar final loss and accuracy, but CNNs exhibited faster initial learning and smoother predictions for shorter horizons. This suggests that CNNs might be better suited for tasks requiring fast convergence and capturing local variations, while DNNs could be beneficial for complex relationships and long-term dependencies. This highlights the importance of careful analysis and evaluation to select the optimal model for specific needs. The authors recommend further analysis based on application-specific requirements to choose the best model.