Time Series Prediction for Measurements of Electric Power Trains

Real-time systems require up-to-date information. Measurement signals in the power train of Electric Vehicles (EVs) are however often received with individual time delays due to the distributed architecture of the power train. Our idea is to compensate the time delays by predicting each signal from the last received value until the present time step. In this work, we evaluate 5 state-of-the-art algorithms and 2 naive methods for time series prediction. We execute all algorithms on real power train data of EVs and compare the results. Our evaluation focuses on run-time and accuracy. All methods achieve a prediction error rate of less than 5 %. As expected, the benchmark naive method is the fastest. Surprisingly, it retrieves comparably accurate results as Exponential Smoothing. BATS and TBATS are the slowest methods. Nevertheless, they achieve the best accuracy, but suffer from outliers. Auto-Regressive Integrated Moving Average (ARIMA) achieves the smallest Mean Absolute Percentage Error (MAPE) and thus the best compromise between outliers and accuracy of all algorithms. Additionally, to further improve the accuracy, we investigate the benefits of combining predictions of different algorithms.