Short-term Load Forecasting Based on Long Short-term Memory Network Combination Algorithm

Short-term power load possesses such features as instability and randomness, so traditional load forecasting methods of present a certain limitation during the modeling. To improve forecasting accuracy, a short-term load forecasting method based on the combination of complementary ensemble empirical mode decomposition (abbr. CEEMD), long short-term memory (abbr. LSTM) and multiple linear regression (abbr. MLR) was proposed. Firstly, by means of CEEMD the power load data was composed into high-frequency component and low-frequency component, then the complex high-frequency component was predicted by Bayesian optimized LSTM neural network, and the periodical low-frequency component was predicted by MLR. Finally, each component was superposed and reconstructed to obtain the final prediction result. In the computing example, in one hand different decomposition methods were compared, and in the other hand different models and the influence of Bayesian parameter adjustment on prediction results were compared. Thus, both reliability and accuracy of the proposed method are verified.