ABSTRACT
Daily records of energy consumption make a time series that can be used for predicting demand expected one or more days in the future. A neural network enhanced with memory neurons and called an MNN, has been developed and tested to make approximate time series predictions. At present, the prediction is done without taking external influences into consideration. In an MNN, both inputs and hidden weights are modified by time delay coefficients. Optimized values for the weights and delay coefficients can be learned incrementally from the systematic components of fluctuations extracted from historical data. Currently, the observed time series is the only source of information being used to predict future values. Later, other series of independent predictive data can be used to expand the MNN model. Network training has successfully demonstrated that up to 6000 iterations over the training set can make a ten-fold reduction in the average prediction error. Repeated presentation of hundreds of training patterns in training a network can lead to systematic convergence and the synthesis of a predictive network. Training of the MNN can make the error of prediction fall to be consistently below 20% when averaged over one year of set-aside test patterns.
