Load prediction of partially observed (digitalized) systems using Machine Learning

This master’s thesis aims to explore the challenges associated with time series datalimitations in short-term energy consumption forecasting (one day ahead) usinginstance-based transfer learning. The study utilized the CityLearn dataset, whichincludes data from nine different buildings over three years, along with weather andcarbon intensity. As a first step, we identify similarity patterns among the buildingsusing hierarchical clustering combined with dynamic time warping (DTW), whichresults in three clusters. The first cluster consisted of buildings 1,2,3 and 4, andthe second cluster consisted of buildings 5,6 and 7, while the third one includedbuildings 8 and 9. The data for each building were combined with carbon intensityand weather data as exogenous features. To prepare the data for training, differentpreprocessing operations were performed, including data cleaning, noise removal,feature engineering, feature extraction, and fine-tuning. A comparison was madebetween the selected models (Long Short Term Memory (LSTM), XGBoost, andRandom Forest) and XGBoost showed the best performance among the test models. Based on this, an instance-based transfer learning method was applied using theXGBoost model. This helps reuse instances or knowledge from source domains thathave a large amount of data to improve generalization in the case of a limited amountof data, as in the case of 20%, 40%, and 60% of the dataset within the same cluster. Acomparison of the model performance was made before and after using the transferlearning. The results show that the average percentage improvement in RMSE forclusters 1, 2, and 3 in this thesis was -36. 95%, 3. 64%, and 0. 73%, respectively. The negative percentage ( -36. 95%) for cluster 1 reveals an increase in RMSE,demonstrating that the proposed model performed worse compared to the baselinefor this cluster. The results highlighted the effectiveness of transfer learning using different machinelearning models in enhancing energy consumption prediction for buildings usinglimited data. KeywordsTransfer Learning, Forecasting, Energy consumption, Time series data, LSTM,XGBoost, RandomForest Regressor, Hierarchical clustering, Dynamic time warping(DTW), Short-term forecasting.