Linear and Neural Network-based Models for Short-Term Heat Load Forecasting

摘要

Successful operation of a district heating system requires optimal scheduling of heating resources to satisfy heating demands. The optimal operation, therefore, requires accurate short-term forecasts of future heat load. In this paper, short-term forecasting of heat load in a district heating system of Ljubljana is presented. Heat load data and weather-related influential variables for five subsequent winter seasons of district heating operation are applied in this study. Various linear models and nonlinear neural network-based forecasting models are developed to forecast the future daily heat load with the forecasting horizon one day ahead. The models are evaluated based on generalization error, obtained on an independent test data set. Results demonstrate the importance of outdoor temperature as the most Important influential variable. Other influential inputs include solar radiation and extracted features denoting population activities (such as day of the week). Comparison of forecasting models reveals good forecasting performance of a linear stepwise regression model (SR) that utilizes only a subset of the most relevant input variables. The operation of the SR model was improved by using neural network (NN) models, and also NN models with a direct linear link (NNLL). The latter showed the overall best forecasting performance, which suggests that NN or the proposed NNLL structures should be considered as forecasting solutions for applied forecasting in district heating markets.