Multistep-ahead prediction of power demand using a sliding window technique and neural networks

This paper presents a new method for prediction of power demand time series using a hybrid algorithm with wavelet decomposition and neural network. The power demand time-series is first decomposed into a certain number levels with discreet wavelet transform and for each individual wavelet sub-series are created neural networks to predict future values. To form the aggregate prediction the individual wavelet sub-series forecasts are recombined using the reconstruction property of wavelet transform. The results are conducted in Matlab software and the performance of this procedure is investigated.