Variable selection for prediction of time series from smart city

Most information and communication technologies systems providing some form of intelligence to future smart cities will more or less use data-based predictive models. Since the amount of data collected increases rapidly, it is becoming crucial to select proper data that are relevant and useful for the specific predictive model. The importance and usefulness of two wrapper feature selection methods is demonstrated here on 23 time series appearing typically in smart city area. Particularly, high dimensionality reduction is achieved without sacrificing the prediction performance for energy consumption, temperature, price and people´s presence prediction. Only for thermal discomfort prediction, high dimensionality reduction causes small increase of mean average prediction error typically less than 1%. Since the two methods are comparable from the dimensionality reduction and prediction performance point of view, sensitivity based pruning is recommended, because of its less computational demands.