Deep Neural Network Based Methodology for Very-Short-Term Residential Load Forecasting

Residential load forecasting has long been a prediction problem due to high uncertainty associated with a single electricity consumer. Predicting residential load demand is important for the efficient operation of utility networks and may consist of the network for the future interactive structure of electricity markets. In the past, many different machine learning algorithms have been applied for load forecasting including deep neural networks for various prediction horizons. However, smart grids and interactive markets will require prediction in very-short term horizons-in the scale of minutes-. This paper seeks to study the use of deep neural networks in very-short-term residential load forecasting. To that end a deep neural network is created and being trained on four different training datasets in order to observe the effect of the training on the network forecast ability. Forecast performance is measured with respect to mean average percentage error on a yearly long 5 min load values of a residential building. The results exhibit that that the deep network is able to make forecasts with MAPE laying in the interva11.1%-1.4% for all the four different training datasets.