Investigation of the Improvement Potential of Heat Load Forecasts in BoFiT

Sammanfattning: Stockholm Exergi is a district heating company which distributes heat to customers in central and suburban Stockholm. The company has five major production plants and as in any energy business is it in the company’s best interest to optimize the production in order to reduce the costs. The optimization plan for the production facilities require heat load forecasts. Consequently does the accuracy of the forecasts have a large impact on how well the heat production can be planned. At Stockholm Exergi the heat load is forecasting performed in the software BoFiT. To facilitate the production planning and ultimately maximize the profits for Stockholm Exergi is the overall purpose of this project to investigate the improvement potential of the heat load predictions in BoFiT. The heat load consist of two parts; space heating and domestic hot water usage. Space heating is primarily weather dependent and the domestic hot water usage is mainly dependent on the social behavior of the inhabitants. Combined does these two create a complex system with stochastic, non-linear and non-stationary characteristics. To forecast such complex systems BoFiT has inbuilt mathematical models. Three different mathematical models in BoFiT have been analyzed; artificial neural network (ANN), regression and SARIMAX. The original model uses an ANN that uses outdoor temperature and previous heat load to forecast the heat load. This study utilize literature, modelling and statistical analysis to identify and evaluate more input parameters in combination with other mathematical models to determine if the model can be improved.   The heat load was thermodynamically and mathematically described based on literature, the BoFiT model was examined and the utilized algorithms was mathematically described. As a basis for the statistical analysis and input parameter implementation was the actual heat load compared to the construction of heat load in BoFiT. The differences was identified and evaluated. BoFiT is inadequate in its method of thermodynamically correct describe the heat load, but due to its ability to learn to forecast based on previous heat load has the model the capacity to generate results with approximately 90-95 % accuracy. To improve the heat load forecasts, this study suggests to shift from calculating the heat load from the difference between the heat imported and the heat exported of a geographical limited area and instead install measurement stations at each customer substation to measure the actual heat demand. The heat load predictions are subjected to uncertainties in the study. These are errors in measurement data, errors in weather forecasts such as outdoor temperature or global irradiance and other forecasts error connected to e.g. supply and return temperature. Another problem with the models is that when more parameters are implemented they can amplify or cancel out other parameters. This is analyzed via a statistical analysis that shows that the multicollinearity is low for the analyzed input parameters. Ergo, the parameters do not affect each other negatively.   The statistical analysis and validation in BoFiT showed that ANN was the best suited mathematical model and that the outdoor temperature and previous heat load should be complemented by the supply and return temperature as input parameters. The results of this study discloses that the original model will not be improved by adding new input parameters and using it to forecast the whole year. The mean absolute average error increased with 18.45 % for the yearly analysis. An improvement is possible if the forecasted year was divided into shorter seasons using different models for each season. The accuracy of the model could then increase in the winter, spring, summer and fall with 0.59 %, 33.74 %, 8.11 % and 4.18 % respectively.   The results from the study was then applied onto another subnetwork to analyze if they were scalable. When only using the best overall model for the whole year the mean absolute error increased even further. This resulted in a model with a mean absolute error increase with 22.47 %. Similarly was the best scenarios for each season applied to the Solna forecast. Here the accuracy of the model decreased in the winter, spring and summer with 8.44 %, 57.07 % and 29.54 % respectively. The fall scenario had an increase in accuracy by 16.22 %.