Using multi-output feedforward neural network with empirical mode decomposition based signal filtering for electricity demand forecasting

For accurate electricity demand forecasting, this paper proposes a novel approach, MFES, that combines a multi-output FFNN (feedforward neural network) with EMD (empirical mode decomposition)-based signal filtering and seasonal adjustment. In electricity demand forecasting, noise signals, caused by various unstable factors, often corrupt demand series. To reduce these noise signals, MFES first uses an EMD-based signal filtering method which is fully data-driven. Secondly, MFES removes the seasonal component from the denoised demand series and models the resultant series using FFNN model with a multi-output strategy. This multi-output strategy can overcome the limitations of common multi-step-ahead forecasting approaches, including error amplification and the neglect of dependency between inputs and outputs. At last, MFES obtains the final prediction by restoring the season indexes back to the FFNN forecasts. Using the half-hour electricity demand series of New South Wales in Australia, this paper demonstrates that the proposed MFES model improves the forecasting accuracy noticeably comparing with existing models.

► Predict the electricity demand of a whole week.
► Use fully data-driven EMD-based signal filtering to reduce the noise of time series.
► Apply the seasonal adjustment method to avoid the seasonal interference.
► Use multi-output strategy to overcome limitations of multi-step-ahead forecasting.
► Construct a multi-input-multi-output FNN forecasting model.