No abstract

Forecasting an hourly heat demand during different periods of district heating network operation is essential to optimize heat production in the CHP plant. The paper presents the heat demand model in the real district heating system with a peak load of 200 MW. The predictive model was developed with the use of the machine learning method based on the historical data. The XGBoost (Extreme Gradient Boosting) algorithm was applied to find the relation between actual heat demand and predictors such as weather data and behavioral parameters like an hour of the day, day of week, and month. The method of model training and evaluating was discussed. The results were assessed by comparing hourly heat demand forecasts with actual values from a measuring system located in the CHP plant. The RMSE and MAPE error for the analysed time period were calculated and then benchmarked with an exponential regression model supplied with ambient air temperature. It was found that the machine learning method allows to obtain more accurate results due to the incorporation of additional predictors. The MAPE and RMSE for the XGBoost model in the day-ahead horizon were 6.9% and 8.7MW, respectively.