A novel time-series probabilistic forecasting method for multi-energy loads

Due to the strong nonlinearity, stochasticity, and high coupling of multi-energy loads, this paper proposes a time-series probabilistic forecasting method based on radial basis function network-based autoregressive with exogenous inputs (RBF-ARX) and Gaussian mixture model (GMM). First, due to the nature of the nonlinear coupling between multi-energy loads, the degree of correlation is quantitatively evaluated by introducing the maximum information coefficient with the optimal feature method. Second, the Fourier transform method is used to categorize the load data into deterministic and stochastic components. For the deterministic component, the RBF-ARX method is used for forecasting, which can realize multiple inputs and multiple outputs. It only requires less training data with faster computation speed than other benchmark methods. For the stochastic component, a GMM with an improved Markov Monte Carlo sampling is used to generate the time-series stochastic component, which can accurately characterize the probabilistic properties. This paper compares the forecasting accuracy and computation time under 4 seasons using the proposed model with benchmark models. Results show that the proposed model can improve forecasting accuracy of multi-energy loads by at least 50 % and the computational efficiency can be improved by at least 30 %, which means a superior performance.