Short-term power grid load forecasting based on VMD-SE-Bilstm-Attention hybrid model

Abstract

Abstract Short-term power load forecasting helps to maintain the equilibrium between the generation and consumption of power, and is also directly related to the supply and demand balance with regard to the power grid and operating costs, thus ensuring the reliable and effective functioning of the electric power system. The traditional single prediction method has poor fitting effect in the face of non-smooth, non-linear fluctuations in the load sequence, resulting in low prediction precision. It is imperative that the aforementioned issues are addressed. This paper puts forth a proposal of a combined prediction method based on the variable modal decomposition combined with the sample entropy and the attention mechanism of the bidirectional long- and short-term memory network. At the beginning of the whole, the VMD is used to decompose the original power sequence into multiple sub-sequences and into components with different frequencies, then the SE values of the original load and modal components are compared and reconstructed into smooth and fluctuating components to reduce the scale of the operation. The different modes are predicted separately by the Bilstm network with consideration of the affecting parameters such as temperature, weather, humidity, and data type, and finally combined with the Attention Mechanism to further explore the correlation within the data to reconstruct and sum the predicted different modal components as the prediction result. Experiments were produced to contrast the prediction results of different models with the actual load of a regional electricity market in the southwestern United States; the models in this paper all achieved better prediction results.