Evaluation of Different Machine Learning Algorithms for Heat Load and Availability Forecasting within the test facility "District LAB"

The focus of this thesis centers on the forecasting of heat demand within district heating systems utilizing a range of machine learning techniques. The District LAB in Kassel serves as a testing facility for the investigation of various aspects of district heating, and the objective of this research is to enhance the performance of district heating systems through the application of data-driven predictive analysis and machine learning methods. The simulation model employed in this research encompasses a district heating system comprising of ten ideal houses, which are modeled using stochastic user behavior, and a plethora of components such as supply and return pipelines, a boiler, a storage facility, a pump, and a weather compensation control system. Two different scenarios were simulated, one without and one with solar thermal collectors, utilizing weather data sets from the German weather service for three distinct years, resulting in a total of six different datasets, inclusive of an average year, an extreme summer, and an extreme winter. After that, these simulation-generated datasets were used for the application of machine learning algorithms. The machine learning algorithms applied to these datasets encompass linear regression, random forest, and long short-term memory (LSTM) models. These algorithms were trained using both univariate and multivariate variables, and the results were compared to identify the most suitable algorithms for predicting heat demand. It was determined that the LSTM univariate model exhibited superior forecasting capabilities in the extreme winter datasets, while linear regression displayed better forecasting abilities in the extreme summer datasets. In an effort to augment the accuracy of the LSTM models, hyperparameter optimization was conducted using the random search method. Although this process proved to be time-consuming and computationally intensive, it resulted in better-fit models. The training of models with these optimized hyperparameters was conducted using a graphics processing unit (GPU) from Google Colab, whereas models without hyperparameters were trained using conventional central processing units (CPU). It was established that when time is not a constraint, LSTM multivariate models displayed superior performance compared to other machine learning (ML) models. But when time is limited, linear regression might be a decent alternative for prediction. The conclusion of this research emphasizes that the combination of data-driven predictive analysis and machine learning techniques can improve the performance of district heating systems by doing a day-ahead prediction of heat demand. The most appropriate machine learning algorithms for predicting heat demand in this context are linear regression and LSTM models. Additionally, it is possible to achieve more accurate predictions by performing hyperparameter optimization on the LSTM models and with sufficient time, LSTM multivariate models outperform all other ML models.