In view of the success of machine learning based prediction algorithms in the recent years, in this study, we have employed a selection of these algorithms on some time series prediction problems in the context of smart grid. We have used real world data from the UCLA campus solar PV panels and parking lots. In the process of applying these algorithms on the Electric Vehicle (EV) charging load prediction problem, two new prediction algorithms have been proposed, namely Modified Pattern Sequence Forecasting (MPSF) and Time Weighted Dot Product Nearest Neighbor (TWDP NN). One of the objectives when predicting the EV charging load is speed of prediction since it is intended to be used in a real time application (smartphone application for EV customers). Using our dataset, TWDP NN decreased the processing time by a third.
As missing data is a significant concern in real world data, the effect of missing values on the prediction quality has been investigated. Six different imputation methods have been applied to compensate for missing values in EV charging data. Based on non-parametric statistical tests, suitable (or unsuitable) imputation methods for each prediction algorithm are recommended.
Forecasting of the Electric Vehicle (EV) charging load can be done based on two different datasets: data from the customer profile (charging record) and data from outlet measurements (station record). We found that charging records provide relatively faster prediction while putting customer privacy at jeopardy. On the other hand, station records provide relatively slower prediction while respecting the customer privacy. In general, both datasets generate comparable prediction error.
Forecasting solar power generation with application on real-time control of energy system has also been investigated. Since predictions are made on every minute for one minute ahead values, the designed system has to be rapidly responsive. This has been pursued by: first, we have solely relied on past values of solar power data (rather than external data), hence lowering the volume of input data; second, the investigated algorithms are capable of generating predictions in less than a second. The results show that kNN and SVR show lower error.
