A short-term electric load forecast method based on improved sequence-to-sequence GRU with adaptive temporal dependence

Accurate and efficient short-term electric load forecast (STLF) is essential for power systems’ reliable and economical operation. The temporal dependence of actual load exhibits dynamics and variability, while the current STLF methods often neglect it, resulting in poor robustness. This paper proposes an STLF method based on an improved sequence-to-sequence gated recurrent unit network (S2S-IGRU) to solve this problem, with a three-step adaptive framework for following dynamic temporal dependency pattern. In the first step, the maximum duration of the temporal dependence is estimated empirically based on the autocorrelation coefficient. It is set as the initial upper limit of the window length to construct the time series samples. In addition, a sine–cosine cycle encoder is used for the periodic categorical inputs. In the second step, an S2S-IGRU initialized with a Chrono initializer is used to forecast the load in upcoming days, which enables a flexible setting of the memory retention time. In the third step, the prediction error changing curve with window length is selected as a fine-tuning criterion to obtain the optimal length of the time window. The results from two real examples justify the proposed method of significantly improving forecast accuracy and adaptability on varied forecast horizons and load datasets with different temporal dynamic patterns.