A Novel ARM-DC AutoConNet for Accurate Long‐Term Time‐Series Forecasting

Time series forecasting is essential in many fields, such as financial analysis, climate forecasting, and energy demand. This ‎study proposes an improved dilated convolutional(DC) AutoConNet architecture to solve the problem of accurate long-term forecasting of complex time series data. The model significantly improves its robustness and generalization ability by ‎integrating multi-scale dilated convolution, layer normalization, and a new adaptive rescaling mechanism(ARM). The ‎main improvement is that while maintaining the efficiency of the original AutoConNet, it effectively solves the overfitting ‎problem and the defect of insufficient capture of temporal dependencies. We evaluated the model performance on 16 ‎standard datasets (including finance, climate, health, etc.), such as M4, M5, ETTh1, ETTh2, ETTm1, and ETTm2. The ‎ARM-DC AutoConNet achieved significant improvements on multiple datasets, especially in the long forecast period, ‎which can significantly reduce the SMAPE index and stabilize the shape value. Comparative experiments demonstrate ‎that the proposed model consistently outperforms or equals the benchmark AutoConNet in terms of error indicators.‎ Furthermore, the proposed model surpasses the AutoConNet Model in error metrics. The most significant improvement is ‎the combination of the adaptive rescaling mechanism and dilated convolution. The improved convolutional architecture ‎provides a feasible solution for reliable long-term prediction and inspires the future development of time series deep learning‎.