Research on Short Term Power Load Forecasting Based on Wavelet and BiLSTM

Abstract

With the developing of big data and artificial intelligence, the application of smart grid has received widespread attention. Specifically, accurate power load forecasting plays an important role in the safety and stability of power system production dispatching. However, traditional load forecasting methods still has some limitations in processing large-scale nonlinear time series data. to accurately predict short-term power loads, a new forecasting method combining BiLSTM and Wavelet decomposition is proposed, named Wavelet-BiLSTM. First, the input time series data is decomposed into different sequences using wavelet decomposition. By comparing the prediction results of different decomposition levels, it is determined that a 2-level decomposition is the most appropriate, resulting in sequences A2, D1, and D2. Next, for each wavelet decomposition coefficient sequence, a separate BiLSTM model is constructed for training and prediction. Next, the prediction results of each Wavelet coefficient series are reconstructed by inverse Wavelet transformation to obtain the final load data prediction value. Experimental results show that the proposed Wavelet-BiLSTM method can improve the prediction precision and accuracy, Therefore, it’s a promising approach for electricity load forecasting tasks.