Unsupervised and supervised machine learning techniques for timeseries aggregation in the design and operation optimization of multi-energy systems

The optimization of the sizes and operation of energy conversion and storage units inmulti-energy systems with time horizons of yearsis strongly dependent on the variable timeseries ofthegrid electricity price.Thus, the optimization could be carried outconsidering representativetimeseriesobtained by unsupervised(e.g., clustering) or supervised (e.g., artificial neural networks)techniques.However, thesetechniqueshave mainly been combinedto improvethe accuracy of the timeseries prediction withoutthenusing these timeseries in an optimization problem.The objective istoevaluatewhether unsupervised techniques can be properly combined with supervised techniques to obtaina representative set oftypicaldaysofthegrid electricity price, the impactof which is assessedontheoptimal total cost of a multi-energy system.This paperproposes a noveland preliminaryhybrid approachthatcombines multi-year clustering and a feedforward Backpropagation artificial Neural Network (BPNN),andthencomparesitwitha state-of-the-art multi-year clustering.Both approaches are fairly applied tothe same “past” dataset (2005-2014), where the multi-year clusteringidentifies clusterscontaining similar timeseries, while the hybrid approachis based ontrainingthe BPNNwith the timeseries labelled according to the representative clusters found bythe multi-year clustering.The multi-year clustering approachfindsa representative set of typical days in the “past” datasetand considers itin a “future” dataset (2015-2020), while the hybrid approachfinds a representative set byclassifyingthe days of the“future”datasetaccording tothetraining of the BPNN in the “past”. Subsequently, astochastic programming modelis used to optimize the design-operation of the system byminimizing itslife cycle (investment and operational)costsinthe“future”dataset, using separately the twodifferentrepresentativesets oftypical daysof the electricity price. The optimal life cycle costs based on the typical days of the multi-year clustering and hybrid approachesshowerrors of 3% and 5%, respectively,compared to“perfect knowledge” solutions based on data really occurred.Preliminary results show the validity of the proposed hybrid approachand point to further improvements.