Machine learning assisted adaptive heat load consumption forecasting in district heating network

Abstract

District heating system often consists of a long, complex network of piping carrying heat from a power plant to the consumers. The supply temperature from the plant is either controlled by the operator from experience or a predefined curve based on the outdoor temperature. An optimized supply temperature which would be lower than the one obtained traditionally would lead to lower heat loss and reduced peak load on the power plant. In this paper, we investigate the machine learning models for heat load forecasting which is a crucial parameter in the optimizing process. Models are generated using supervised machine learning algorithms: Linear models (Linear Regression, Ridge and Gaussian Process Regressor), Random Forest Regressor, Support Vector Machine (SVM) and Long Short-Term Memory (LSTM) recurrent neural network (RNN). Data-driven models are used extensively in the literature to predict heat load prediction based on the weather and the time effect on a fixed training set, however, in this study, we model the heat load in the network in real-time scenarios i.e., adaptive training and forecasting. The model is adaptively updated as well as the training of the machine learning model in real time. It provides a “plug-and-play” solution for real-time prediction without significant pre-tuning requirements. The results of all the models are compared with various time horizons i.e., 6 hrs, 10 hrs, 24 hrs and 1 week, using the district heating data obtained for the city of Vasteras in Sweden. The performance of the prediction algorithms is evaluated using Mean Absolute Percentage Error (MAPE) and Root Mean Squared Error (RMSE). An algorithm with the best accuracy is selected based on the performance comparison. Also, models suitable for short-term and long-term forecasting are discussed towards the end of the article