| Sumario: | Time series forecasts are essential for understanding and anticipating patterns in data that vary over time. These predictions apply across a variety of fields, from meteorology, where they are used to forecast future weather conditions, to the financial market, to anticipate movements in stocks and currencies. This thesis details the innovation brought about by the integration of dense networks, aimed at improving both the modeling and the accuracy of predictions. Surpassing the SCINet model, which was already recognized for its good results in univariate and multivariate time series, the study introduces DESCINet. This new model resolves issues identified in SCINet, particularly those arising from the use of downsampling, a technique that, despite being useful, could lead to the loss of critical information and heavily depended on fine-tuning of hyperparameters. Furthermore, the thesis addresses the difficulty SCINet had in maintaining accuracy in long-term forecasts due to its limited ability to capture complex patterns across various temporal scales. DESCINet, with its approach of dense residual connections, promises to overcome these barriers, preserving detailed information and enhancing the ability to model complex temporal dependencies. This innovative approach allows the model to maintain consistent performance across extended forecasting horizons. The practical application of DESCINet was tested on a wide range of data sets, such as ETT, Weather, Electricity, Illness, Traffic, and Exchange Rate. In all these cases, DESCINet demonstrated superiority over SCINet, validating its efficacy in varied and complex contexts. The selection of these datasets illustrates the diversity of challenges inherent in time series forecasting and highlights the adaptability and robustness of DESCINet. This study contributes to the field of time series by exploring the yet untapped potential of dense networks. Integrating these networks into time series forecasting models paves the way for significant advancements, both academically and in practical applications. The proposition of DESCINet indicates a promising direction for future research, suggesting that overcoming current limitations in time series forecasting is within reach. In conclusion, this thesis offers a better understanding of the impact of dense connections on time series forecasting, encouraging the scientific community to investigate DESCINet more deeply. The work is expected to stimulate ongoing research in this area, laying the groundwork for new innovations and practices that enhance time series modeling, making it more effective and precise.
|
|---|
