Electricity load forecasting by randomized neural networks

• Main Author: Gao, Ruobin
• Other Authors: Yuen Kum Fai
• Format: Thesis-Doctor of Philosophy
• Language: English
• Published: Nanyang Technological University 2022
• Subjects: Engineering::Electrical and electronic engineering::Electric power
• Online Access: https://hdl.handle.net/10356/161497
• Institution: Nanyang Technological University

Recently, a hot research topic has been time series forecasting via randomized neural networks and its application in the energy system. Electricity load is a representative time series from the energy system. Forecasting the electricity load accurately benefits electric power system planning for maintenance and construction. As a result, developing novel and accurate forecasting models for short-term load is significant. Among the vast energy forecasting models, a huge branch is the hybrid models with signal decomposition blocks. However, most literature does not implement the decomposition properly. A walk-forward decomposition is proposed, which overcomes the data leakage problem—such decomposition functions by processing the observations in the moving window. The latest observations are fed into the machine learning model for forecasting purposes. We adopt the random vector functional link (RVFL) network, whose hidden features are randomly initialized and fixed during training. Finally, the proposed model consists of two blocks, the walk-forward decomposition and the RVFL. Deep learning, the emerging technique of deep representation learning, is investigated in this thesis by combining it with shallow randomized neural networks. Sections 4.3 and 4.4 in chapter 4 propose an ensemble deep RVFL, and signal decomposition is combined with it to increase the accuracy further. Unlike classical deep learning, the proposed model constructs an independent output layer for each hidden layer. Finally, ensemble learning is utilized to combine their outputs. Then, chapter 6 extends this architecture to randomized recurrent neural networks, which are echo state networks (ESN). A heterogeneous ensemble deep ESN is proposed by combining redundant wavelet transformation. An independent ensemble deep ESN is constructed on each scale of the wavelet transformation, and an aggregation of all forecasts is the output. This thesis contributes to the literature and industry from the following perspectives. First, this thesis contributes to the literature about decomposition-based forecasting by proposing a causal and novel decomposition scheme. Second, three novel decomposition-based forecasting models are proposed. Third, the shallow randomized neural networks from the literature are extended to deep architectures for forecasting. Besides, the thesis also contributes to the industry significantly. First, the improved forecasting accuracy guarantees the balance between energy supply and demand. Second, accurate load forecasting also reduces energy generation costs. The accurate forecasts can increase the energy efficiency and profits for the energy generating and distributing companies. Finally, the related companies can plan because of the accurate estimations of future consumption.