A short-term load forecasting model for demand response applications

Abstract

This paper discusses a new algorithm and defines the functionality required for developing a short-term load-forecasting module for demand response applications. Feedforward artificial neural network (ANN) algorithms are used to provide high forecasting performance when dealing with nonlinear and multivariate problems involving large datasets. The approach is thus suitable for short-term load prediction for disaggregated sites to optimize the demand response process when the data relating to the operating regime or load characteristics of the individual devices and loads connected are unavailable. A detailed description of the relevant external data needed for the forecast is explained. In particular, the algorithm considers weather data for the corresponding time period. The model is tested on data from actual ground source heat pump (GSHP) and heating, ventilation and air conditioning (HVAC) loads of various non-residential buildings at several real sites in the United Kingdom (U.K.). The sensitivity of the parameters of the algorithm, including the number of hidden layers used, is also researched. The proposed algorithm is tested against a linear regression and proves to outperform the latter in all cases. The performance of the algorithm is quantitatively assessed using mean absolute per cent error and mean absolute error metrics. Further analysis plots a comparison of actual and forecasted loads and R-values to determine forecast accuracy.