Monthly electricity demand forecasting using empirical mode decomposition-based state space model

Guaranteeing stable electricity demand forecasting is paramount for the conservation of material resources. However, because electricity consumption data are often made up of complex and unstable series, it is very hard for a simple single method to always obtain accurate predictions. To improve electricity demand forecasting robustness and accuracy, a hybrid empirical mode decomposition and state space model are proposed, for which the empirical mode decomposition is applied to decompose the total time series (noise filtering), and the state space model is employed to forecast every sub-series (feature extraction), with the state space model parameters being optimized using maximum likelihood via a Kalman filter. Compared with autoregressive integrated moving average model and artificial neural networks, the proposed model had more stable and accurate forecasting. This method could be broadly applied to not only forecast electricity demand, being a key step for developing electricity generation plans and formulating energy policy, but also forecast any similar time series data with noise and substantive latent features, making a new step toward solving such a problem.