Multivariate Time series Forecasting with applied Machine Learning on Electrical signals from High-Voltage Direct Current Equipment - Valve Cooling System

Abstract: In a sustainable society, utilizing intermittent renewable power plants is an important building block for achieving green power production. However, the power production from these sources, e.g., wind farms and solar farms, are often located far away from the place of power consumption, and the electricity generation is affected by the weather conditions in the area. Therefore, there is a challenge in balancing power production and consumption with these sources. The HVDC (High-Voltage Direct Current) technology can be used to efficiently transport electricity over long distances and is a key concept in the utilization of renewable energy sources. However, the HVDC systems are sensitive to environmental effects such as elevated or dropping ambient temperatures, which can cause a forced stop in the system, e.g., when the remaining cooling capacity is low. Therefore, the HVDC systems are built to have a high redundancy to maintain a secure power transmission during seasonal changes.  This thesis aimed to create a forecasting model with applied machine learning that could trend the remaining cooling capacity in an HVDC system, to stay aware of how much remaining cooling capacity there is at different seasons. This can be used to optimize the power transmission during seasons when there is a surplus of cooling capacity. The machine learning pipelines were constructed in Python utilizing Hitachi Energy’s PGML (Power Grid Machine Learning) platform. Two different forecasting models were used: LSTM (Long Short-Term Memory) and XGBoost (eXtreme Gradient Boosting). The models were trained to make a five hour ahead multistep prediction and were validated with several evaluation metrics. The best performing model was the XGBoost model, therefore it was chosen as the final model and was tested on a hold-out data set to estimate the general performance. The final model performed well on the hold-out data set, based on the scores from evaluation metrics. Residual diagnostics were used to improve the models during training and to evaluate the final model. At the end of the discussion in Chapter 5 future improvements were suggested.