PSR-TSTrans: a Novel Two-stage Predictor Based on Phase Space Reconstruction for Ultra-short-term Power Load Forecasting

USTLF plays a pivotal role in contemporary power systems, particularly in the realm of online forecasting for power system operations. In order to solve the essential problem of external information lacking and for better mining of non-stationary features of electricity loads to improve forecasting accuracy in autoregressive forecasting scenarios, a novel two-stage deep learning model based on the theory of phase space reconstruction(PSR) is proposed, in which Conv-LSTM stacks are implemented as the temporal feature extractor for the non-stationary dynamic characteristics of chaotic system in the first stage, and Transformer encoder stacks are employed to leverage the spatial features of the K-nearest neighbors in the phase space for adaptive correction in the subsequnet stage. The proposed model is named as phase space reconstruction based two stage Transformer, PSR-TSTrans. A real world dataset provided by the Energy Market Company in Singapore(EMC) is used as the benchmark, six classical works for time series forecasting along with the variant of PSR-TSTrans are selected as the baselines for comparison, and the results show that our proposed PSR-TSTrans model can achieve a MAPE score of 0.475% in average for ultra-short-term load prediction, which is much more accurate than the baselines, and the ablation experiments illustrate that the two-stage prediction approach yields additional enhancements in prediction accuracy when compared to the traditional one-stage prediction method commonly utilized in chaotic time series forecasting, confirming the effectiveness of the method.