Hierarchical Load Forecasting Considering the Penetration of Distributed Energy Resources

: (Alternative Format Thesis)

Abstract

The increasing adoptions of distributed energy resources (DERs) have shifted the traditional energy system to a decentralised arrangement and passive consumers to active prosumers. Traditional load forecasting that focuses on the system level has to move towards the grid edge. Hierarchical load forecasting (HLF) has been recognised as a promising solution to reveal granular load patterns and customer diversity through a two-stage process: 1) the hierarchical model that organises customers into groups and optimises the group portfolio based on the similarities of their natural attributes or consumption behaviours (i.e. customer grouping/clustering); 2) the most appropriate forecasting model is then assigned to each group to respect their distinctive characteristics.

Due to the lack of a comprehensive understanding regarding the impact of customer diversity to the forecasting performance, the traditional methods connect these two stages in a sequential way that could be considered as an open-loop design. The major drawback of this open-loop design is revealed through the investigation into the relationship between customer diversity and the distribution of forecasting errors, demonstrating that the existing hierarchical models to minimise the within cluster variance do not lead to a unique forecasting model that delivers the best forecasting result for the group.

This thesis addresses this challenge by aligning the optimisation objectives through a closed-loop design to identify optimal customer portfolios from the perspective of improving forecasting accuracy. The novelty of the proposed methodology lies in the introduction of a feedback mechanism to return the forecasting error as a signal for the hierarchical model to optimise both the hierarchical model and forecasting model. In this way, the hierarchical model will identify the most compatible customer group in terms of reducing the overall errors in HLF to provide additional (or alternative) criterion to cluster customers.

This thesis develops two novel models to implement the proposed methodology:

i) A Closed-Loop Clustering (CLC) algorithm based on statistical models: it uses the forecasting error signal from the feedback mechanism as the clustering criterion to iteratively update the hierarchical model. In this way, the hierarchical model is enhanced by re-assigning the cluster membership and the forecasting models are updated correspondingly, thereby achieving the co-optimisation of hierarchical model and forecasting model.

ii) A Stacked Model Selection Network (SMSN) based on artificial intelligence: a series of forecasting models are stacked as neurons which are updated in a competitive manner and form the space for the model selection of samples. The forecasting error metric from the feedback mechanism not only serves as the loss function for the update of forecasting models but also the criterion for the winning neurons so that each neuron collects its most compatible sample group. The feedback mechanism enables the interaction between the forecasting system and the model selection system and hence can automatically select the best forecasting model for a mixed dataset.

The potential benefits of the proposed methodology are demonstrated through a local energy market (LEM), which aims to enhance the local absorption of DERs. A major obstacle to maximise the value of LEM is energy forecasting due to its high level of diversities and granularities. By employing the proposed methods in local energy trading, the forecasting improvement brought by optimal prosumer grouping has increased the local PV absorption by 9.03% (CLC) and 12.35% (SMSN) and the revenue of PV owners by 6.53% (CLC) and 9.48% (SMSN) on average compared with representative hierarchical forecasting methods.

Date of Award	16 Jun 2021
Original language	English
Awarding Institution	University of Bath
Supervisor	Furong Li (Supervisor) & Chenghong Gu (Supervisor)