A COMMITTEE OF COIF 3 WAVELET RECURRENT NEURAL NETWORKS FOR ONE-DAY-AHEAD ELECTRICAL POWER LOAD DEMAND PREDICTION

摘要

The problem of predicting one-day-ahead electrical power load demand is solved. Various optimal predictors for committee of recurrent neural networks are designed on each part of the coif 3 wavelet-transformed data to achieve final prediction precisely by adding the individual forecast of optimal predictors on all components. Feasibility of compactly supported coif 3 wavelet with suitable number of decomposition levels are investigated to choose the suitable level of resolution for different seasonal load series. The efficacy of the estimated models are evaluated over different scales such as by partitioning of the randomly chosen data set for ensuring that the proposed technique is neither biased for specific data nor for any partitioning scheme for obtaining good results allowing training patterns for different periods of time, different number of epochs and different number of retraining. The capability of the proposed technique is justified by reasonably low Mean Absolute Percentage Error (MAPE) for various weather patterns. The reliability and consistency in prediction by the adopted technique is justified even in the presence of controlled uniform and Gaussian noise in input channels.