Merus Power is participating in Energy ECS project to create secure and smart energy solutions for mobility in the future. The project has six use cases and Merus Power is leading in UC 3 for smart mobility with Emobility for charging and discharging e-vehicles. This work was focused on researching and implementing machine learning methods for predicting energy consumption. The goal was to find optimal deep learning architectures, add different features and fine tune hyperparameters to discover efficient solutions.

Information was gathered from different web pages from professionals in the field of AI. Online learning platforms which offer books on the topic of machine learning and deep learning were used. Also, thesis supervisors were sharing information on the topic of machine learning.

Working method was implemented with CRISP-DM which has clear steps for understanding the goal, gathering the data, and processing it, then proceeding into modelling and evaluating the results. This methodology is agile and has possibility to iterate between steps.

Results were that to get proper predictions the data needs to be good quality, properly cleaned and processed. RNN were better choice than statistical models. GRU and LSTM had almost the same results and after temperature was added as feature the results improved slightly. Having 2-3 layers and 30-50 units got least error in the predictions meaning less layers and units were better. Learning rate of 1e-3 was optimal.

As conclusion RNN models are effective to predict time series with single or multiple features. Results could be improved by building the models with TensorFlow low level components and experimenting with different features and using moving average to get more precision for each time step prediction.