Regional Short-term Load Forecasting Based on Cluster Identification and Convolutional Neural Network-Bi-directional Long Short-term Memory-temporal Pattern Attention

In this paper, a predictive method based on cluster identification and convolutional neural network-bi-directional long short-term memory-temporal pattern attention (CNN-BiLSTM-TPA) is proposed to solve the problem of excessive input features and strong load periodicity in regional short-term power load forecasting. Firstly, load nodes within the region are identified as clusters based on second-order clustering algorithm with consideration of power consumption mode and weather as the influence factors. And then, the representative features are selected from each cluster as inputs of the deep learning model, which can not only reduce the input feature dimension and decrease the computational complexity, but also comprehensively consider the overall characteristics of the prediction region to improve the prediction accuracy. Thereafter, aiming at the strong load periodicity of regional power load, the CNN-BiLSTM-TPA model is trained and applied for prediction, extracting the bi-directional information from the input data to generate the hidden state matrix and weighing the important features of the hidden state matrix, while capturing the bi-directional time series information on multiple time steps for prediction. Finally, the effectiveness of proposed method is verified using the actual load data in California, USA.