Application of Long-Short Term Memory Network and its Variants in Short-term Power Load Time Series Forecasting

Accurate and effective short-term power load forecasting (STLF) is very important for power systems Considering the temporal and non-linear characteristics of power load, this paper studies the application of standard long-short term memory (LSTM) network and its two typical variants, the Gated Recurrent Unit (GRU) and the Just Another NETwork (JANET) in STLF. So as to evaluate performance of the LSTMs, this paper compares three networks from prediction accuracy and training time. Furthermore, the hyperparameters of all LSTMs and the selection of training samples and inputs in the standard LSTM are discussed. The results of two actual examples show the LSTMs are all able to achieve accuracy of over 92% for the day-ahead hourly load prediction, but the standard LSTM and GRU have higher accuracy (over 94%), and the JANET has higher efficiency. Also, the optimal input window size is closely related to the power load period, and the prediction accuracy of LSTMs for STLF can be significantly improved through filtering training samples based on day-type or adding the loads of previous day to the inputs.