GC-DAWMAR: A Global-Local Framework for Long-Term Time Series Forecasting

Current Long-term Time Series Forecasting (LTSF) approa-ches struggle to capture long-range correlations of prolonged time series. They lack efficient solutions for distribution shift, excessive stationarization, and overfitting caused by training noise. Global convolution and de-stationary autocorrelation are used in GC-DAWMAR, a long-term time series forecasting approach, to address these issues. The global-local architecture maintains translational invariance while capturing inter-subsequence relationships. The de-stationary autocorrelation technique prevents excessive stationarization, while exponential moving average optimization regularization reduces training overfitting. On three real datasets, the suggested LTSF technique outperforms baseline algorithms in prediction accuracy.