Parallel multi-step ahead power demand forecasting through NAR neural networks

We consider the problem of power demand forecasting in residential micro-grids. Previous approaches rely on ARMA models and, recently, on neural network architectures that are however solely used to perform one-step ahead predictions. Here, we propose an original forecasting technique based on non-linear, autoregressive (NAR) neural networks. Our architecture allows for parallel and efficient training and is also lightweight at runtime. Time series from real power demand traces are used to validate our technique, assessing its superiority with respect to state-of-the-art ARMA and ARIMA estimators. Besides smaller prediction errors in the mean and the variance, the proposed NAR architecture provides reliable indications of rises in the power demand even when predictions are generated for a time span of 2 hours. In this case, standard ARMA and ARIMA models entirely fail.