Online training algorithms based single multiplicative neuron model for energy consumption forecasting

• Published in: Energy (Oxford) Vol. 59; pp. 126 - 132
• Main Authors: Wu, Xuedong, Mao, Jianxu, Du, Zhaoping, Chang, Yanchao
• Format: Journal Article
• Language: English
• Published: Kidlington Elsevier Ltd 15-09-2013; Elsevier
• Subjects: Algorithms; Applied sciences; case studies; data collection; Economic data; Electric energy; Energy; Energy consumption; Energy consumption forecasting; Energy economics; equations; Exact sciences and technology; filters; Forecasting; General, economic and professional studies; Iterated nonlinear filters; Iterative algorithms; Mathematical analysis; Mathematical models; Methodology. Modelling; Neurons; Nonlinear filters; rolling; Single multiplicative neuron model; Small dataset and nonlinearity; Asia; United States; North America; China; America; Single multiplicative neuron model; Iterated nonlinear filters; Energy consumption forecasting; Small dataset and nonlinearity; Performance evaluation; Energy consumption; Nonlinearity; Modeling; Forecasting; Kalman filter
• ISSN: 0360-5442
• DOI: 10.1016/j.energy.2013.06.068

Although traditional approaches can yield accurate forecasting results of energy consumption, they may suffer from several limitations such as the need for large dataset and the linear assumption. Two novel hybrid dynamic approaches, which are based on the SMN (single multiplicative neuron) model and the iterated nonlinear filters, have been proposed for forecasting energy consumption with small dataset and nonlinearity in our study. The forecasting models are established by using the weights and the biases of SMN model to present the state vector and the output of SMN model to present the observation equation, and the input vector to the SMN model is composed of the known energy consumption values with a rolling mechanism. The SMN model has advantages of better approximation capabilities, simpler network structures and faster learning algorithms. The nonlinear filters can deal with additive noises and can update model parameters when a new observation data arrives due to their iterative algorithm structure. Two case studies of energy consumption have been used to demonstrate the reliability of the proposed models, and the experimental results have indicated that the proposed approaches outperform existing models in forecasting energy consumption.
•Two novel hybrid dynamic models have been proposed.•The proposed models have better approximation capabilities, simpler network structures and faster learning algorithms.•Using only few historical data, the proposed models can obtain excellent forecast accuracy.•Compared with alternative methods, the proposed approaches have shown better performance.•Comparison and analysis between the proposed two novel models for two case studies of energy consumption are provided.