A Short-Term Power Load Forecasting Method Taking into Account Similar Days

The power load forecasting, which is an important part of energy management systems, tends to suffer from nonstationary and noisy data. Firstly, the historical load data are categorized into four seasons by seasonal characteristics, and the raw data for each season is decomposed and denoised once using the variational modal decomposition (VMD) algorithm to remove high-frequency noise. Then, similar days are filtered as the initial dataset using the combined dimensionality reduction clustering algorithm of <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$t$</tex>-distributed stochastic neighbor embedding (<tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$t$</tex>-SNE) and K-means++ on the denoised data of each season. Next, the complex nonstationary time series are further decomposed into several more stationary and frequency-concentrated intrinsic mode functions (IMFs) components by secondary decomposition of the similar day dataset for each season using complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN). The improved hippopotamus optimization (IHO) algorithm is then used to search for the optimal hyperparameters of the bidirectional long short-term memory (BiLSTM) network, which is applied to predict each IMF component individually. Finally, the predictions of each component are fused to obtain the final forecasting. Compared with seven other forecasting methods, the proposed method achieves a forecasting accuracy exceeding 98% in each season, demonstrating superior performance and strong practical applicability.