Alleviating distribution shift and mining hidden temporal variations for ultra-short-term wind power forecasting

Abstract

Randomness and non-stationarity are common challenges in wind power forecasting (WPF). Many studies focus on randomness but usually ignore the non-stationarity which leads to distribution shift and affects prediction accuracy. To address the distribution shift problem, an alleviating distribution shift using the recent difference characterization (Dish-RDC) method is proposed as a general neural paradigm for WPF. Dish-RDC categorizes the distribution shift into intra-space and inter-space shifts. By employing the RDC, the method facilitates the mapping of input sequences to learnable distribution coefficients that better estimate the distribution. Furthermore, real-world time series often exhibit multi-periodicity, yet existing models face limitations in capturing this temporal variation. To address this issue, our research introduces the Temporal 2D-Variation Model (TimesNet) in WPF. This innovative model extends time variation analysis into a 2D space based on multi-periodicity. By using 2D kernels to model these variations, TimesNet can effectively incorporate advanced computer vision techniques into WPF. Combining these approaches, we developed Dish-RDC-TimesNet, a hybrid model. Experiments show it reduced mean absolute error (MAE) by 47.10 % and 20.63 % on two datasets compared to benchmark models. Moreover, integrating Dish-RDC with benchmark models decreased MAE by 39.79 % and 17.85 % on these datasets.