Enhancing Short-Term Power Load Forecasting for Industrial and Commercial Buildings: A Hybrid Approach Using TimeGAN, CNN, and LSTM
Yushan Liu, Zhouchi Liang, Xiao Li
Abstract
The application of smart meters was delayed, leading to sparse power load data collection in industrial and commercial buildings, often encompassing only days to a few months of data. In contrast, deep learning models necessitate extensive data sets, spanning several years. To bridge this data deficit, this paper proposes a hybrid forecasting method combining Time-series Generation Adversarial Network (TimeGAN) with a Convolutional Neural Network (CNN)-enhanced Long Short-Term Memory (LSTM) neural network. Initially, the scarce data set is expanded using synthetic data derived from TimeGAN. Subsequently, the comprehensive data undergoes CNN filtering, optimizing the information extraction and expediting the forecasting network. The extracted information is then channeled into LSTM network for load forecasting. A case study is carried out using two-month power load data from four different industrial and commercial buildings types, underpins this methodology. Comparative analysis reveals that the proposed model effectively improves short-term power load forecasting accuracy.