Short-Term Load Forecasting Reliability in Power Plant of Cyber-Physical Energy System Considering Adaptive Denoising

• Published in: IEEE systems journal Vol. 17; no. 4; pp. 5183 - 5194
• Main Authors: Ding, Dong, Li, Junhuai, Wang, Huaijun, Wang, Kan
• Format: Journal Article
• Language: English
• Published: New York IEEE 01.12.2023; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: 2-D deep temporal convolutional network; Accuracy; Adaptive denoising; Adaptive systems; Automatic control; Correlation coefficients; cyber-physical energy systems; Cyber-physical systems; Electric power demand; energy management; Feature extraction; Forecasting; Impact loads; Load forecasting; Load modeling; Mathematical models; Noise reduction; Plant reliability; Power plants; Predictive models; Reliability; short term load forecasting; Smart grid
• DOI: 10.1109/JSYST.2023.3310548
• ISSN: 1932-8184

Cyber-physical energy systems (CPES) are a crucial component of smart grids (SGs), and as such, they represent a specialized subset of cyber-physical systems. CPES provides essential services for pricing decisions and automatic generation control through short-term load forecasting (STLF), making the accuracy of STLF critical to optimizing their operation. However, due to the numerous communication devices installed within CPES, data collection is often subject to various factors that could negatively impact load forecasting accuracy. To improve the accuracy of STLF, this article proposes a reliable method that combines an adaptive denoising technique, a 2-D deep temporal convolutional network (TDeepTCN), and a multidimensional input structure bidirectional long short-term memory-attention (MBiLSTM-attention) network. First, an adaptive approach that combines Pearson correlation coefficient and complete ensemble empirical mode decomposition with adaptive noise is utilized to effectively identify raw load series contaminated by noise and reconstruct them. Then, a TDeepTCN model is constructed using TCN to simultaneously capture and fuse both local and long-term temporal features from multiple load series. Finally, MBiLSTM-attention is employed for accurate forecasting to achieve feature processing for multidimensional depth features. Eventually, compared to existing models, our proposed model achieves the most accurate forecasting results with a mean absolute percentage error rate of only 3.98% and 4.12%, respectively, in both regions.