Training data selection using information entropy: Application to heating load modeling of rural residence in northern China

Abstract

The selection of input variables and their amount has been an important issue in big data load forecasting. Taking heating load forecasting as an example, this paper proposed a method for data filtering based on information entropy. First, the heating data from an air source heat pump system adopted by a rural residence in northern China were employed. Moreover, the training data were classified based on linear or nonlinear variations of outdoor temperature and its changing ranges, while the validation data included three different types of weather conditions, namely, cold, cool, and mild. Then, the information entropy under 2-h, 4-h, 6-h and 8-h training window was quantified to be 1.811, 1.839, 1.877 and 1.856, respectively. For the employed rural residence, an equivalent three-resistance and two-capacity model was established to validate the effectiveness of the training window. Using the derived optimal thermal resistance and capacity, the various selection of outdoor temperature variation trend and range were compared and optimized. Results showed that 6 h of training data had the maximum information entropy and the most abundant information, the minimum errors between actual and forecasting data were observed under 6 h of training data, linear change, and lower outdoor temperature. The mean absolute percentage errors for the load forecasting of three typical days were 5.63%, 8.46%, and 12.10%, respectively.