AMIDiNe Work Package 2: Providing load understanding and insights

The thorough analysis of consumption readings as provided on a per MPAN basis will allow us to develop understanding of the different behaviour that is visible only through users energy consumption. New and state-of-the-art unsupervised machine learning methods (clustering algorithms) to group and generate insights from daily profiles and time series will be investigated. This information will be directly used as feedback to WP3 in three ways: 1) to create specific models of different segmented groups of users where the segmentation is based on different energy consumption features and not just metadata as input to the hierarchical forecasting framework and 2) to characterise points of aggregation within the network, such as substations, aggregating the typical consumption profiles for clusters and 3) to help in the creation of possible scenarios where LCT are included. Key tasks for WP2 are:

Disaggregating a common MPAN – we will investigate the effectiveness of a range of approaches for disaggregating the load profile obtained directly at the MPAN, invoked by ‘private wire’ supply arrangements. Blind source separation approaches (e.g. Empirical Mode Decomposition, Probabilistic Time Frequency Masking or Independent Component Analysis) will be used to disaggregate underlying profiles that cannot be observed through the common MPAN. By providing a higher order classification, as opposed to a highly detailed demand breakdown, we will separate the profile into generation/storage and consumption classes and inform the capabilities for local and regional network operation.

Deriving commonality amongst MPAN behaviours to increase the visibility of customers who adopt new technologies such as PV panels, batteries and EV to the network. Dirichlet Process based clustering methods will be employed to describe consumer demand profiles as groups of sub-populations that can be subsequently treated with operation algorithms for network performance. This will build on our experience in the analysing of domestic customers and non-domestic customers but now through the lens of how flexibility will affect the visibility of different technologies within the data collected.

Regional Load Growth: Demand profile migration, dynamics and prediction will be considered in terms of the demand-side response and behavioural change motivations. By adopting a ‘mover-stayer’ approach various estimators of external perturbations such as a cold weather events, clustered EV charging or internally social or domestic driven changes such as adoption of storage or low carbon heating will be investigated. The group dynamic models responses to these drivers can be captured in terms of stability characteristics and can be formally assessed using established tools from other domains such as bifurcation analysis.