Using matrix factorisation for the prediction of electrical quantities

The prediction task is attracting more and more attention among the power system community. Accurate predictions of electrical quantities up to a few hours ahead (e.g. renewable production, electrical load, etc.) are for instance crucial for Distribution System Operators (DSOs) to operate their network in presence of a high share of renewables, or for energy producers to maximize their profits by optimizing their portfolio management. In the literature, statistical approaches are usually proposed to predict electrical quantities. In the present paper, we present a novel method based on matrix factorization. Our approach is inspired by the literature on data mining and knowledge discovery and the methodologies involved in recommender systems. The idea is to transpose the problem of predicting ratings in a recommender system to a problem of forecasting electrical quantities in a power system. Preliminary results on a real wind speed dataset tend to show that the matrix factorization model provides similar results than ARIMA models in terms of accuracy (MAE and RMSE). Our approach is nevertheless highly scalable and can deal with noisy data (e.g. missing data).