Abstract
Heuristic or clustering based time series aggregation methods are often used to reduce temporal complexity of energy system models by selecting representative days. However, these methods potentially neglect relevant information of time series (e.g., distribution parameters). To identify relevant time series parameters, feature selection algorithms can be applied. The present research contributes by (a) developing a new feature selection approach based on clustering, nested modeling and regression (CNR) which is designed for applications requiring high selectivity and using different data sets, (b) comparing and evaluating CNR with feature selection methods available from the literature (e.g., LASSO) and (c) identifying relevant information of the time series applied in energy system models, in particular those of demand, photovoltaic and wind. Results show that CNR achieves on average up to 101 % lower mean absolute errors when methods are directly compared. Thus, CNR better identifies relevant information when the number of selected features is restricted. The disadvantage of CNR, however, is its high computational effort. A potential remedy to counter this is the combination with another method (e.g., as pre-feature selection). In terms of relevant information, energy systems including photovoltaic are mainly characterized by the correlation between demand and photovoltaic time series as well as the range and the 35 % quantile of demand. When energy systems include wind power, the minimum and mean of wind as well as the correlation between demand and wind time series are relevant characteristics. The implications of these findings are discussed.