UniMATS: A Unified Time Series Forecasting Model With Multi-Dimensional Attention Structure for Power Systems

Due to the intrinsic complexity of time series forecasting within power systems, artificial intelligence has emerged as a promising pathway for predictive analytics. Although time series data from power systems inherently exhibit consistent characteristics across spatial, temporal, and covariate dimensions, current forecasting methods remain scenario-specific and fail to develop unified models. This limitation stems from incomplete utilization of multi-dimensional features, resulting in suboptimal data exploitation. To this end, this paper initially constructs a unified tensor representation for time series data of power systems, encapsulating various forecasting tasks within an integrated forecasting architecture. Following this, a unified model with multi-dimensional attention structure for time series forecasting (UniMATS) is proposed, which is capable of integrated feature extraction across all data dimensions simultaneously. Evaluated on three public datasets of wind power, photovoltaic, and load, UniMATS achieves RMSE reductions of 9.8%, 6.9%, and 14.6% respectively compared to the second-best models in day-ahead forecasting tasks. The results of the case study underscore the superior performance and the efficient model structure of UniMATS over existing benchmarks, affirming its general capabilities for time series forecasting in power systems. The general predictive capability and the modular design of UniMATS enable the development of a foundational large model for time series forecasting in power systems based on UniMATS blocks.