Retrieval, Ranking and Recommendation

Part 5 established what groups of customers can and cannot tell a Distribution System Operator (DSO). Archetypes of the shape of demand carried almost none of a customer’s network risk; the transformer explained most of the voltage risk, the household’s level of demand most of the thermal risk, and the regimes of violation indicated what kind of remedy a transformer needs. The decisions a DSO takes, however, concern individual customers and transformers, and they come in a fixed order. A connection request must be answered, often before any study of it has been run; the transformers and customers that need attention must be put in an order, because there is never the capacity to examine all of them; a remedy must be chosen for each, from operational settings to a larger transformer; and, once the network is in operation, the violations that occur at the same time must be put in an order for the operators who respond to them.

These are problems of retrieval, ranking and recommendation, which information retrieval, recommender systems, case-based reasoning and causal inference have studied for decades. This part borrows that theory, adapts it to the Low Voltage (LV) network, and evaluates every method against simulated network outcomes. The simulator supplies what recommender systems almost never have, the outcome of every alternative action for every case, so that a recommendation can be scored exactly rather than estimated from logged behaviour.

Chapter 1 states the theory in the book’s notation and illustrates each method on a synthetic case base whose outcomes are known for every mitigation measure. Chapter 2 applies retrieval and ranking to the AusNet networks. It answers connection requests by retrieval of similar cases, with decisions whose rate of wrong approvals is bounded by conformal risk control; it ranks transformers and customers by their margin, by the onset and severity of PV violations, and by risk under uncertain adoption; it tests whether a ranking learned on one network transfers to a second; and it shows the roles that the clusters of Part 5 take in this chain, as strata for calibration, as a basis for selecting the transformers to study, and as the groups to which programmes are addressed. Chapter 3 recommends a mitigation measure for each transformer. It compares the export limit, managed charging of electric vehicles, a change of transformer tap and reinforcement by their effect on each kind of violation and by their cost, evaluates recommenders from a rule on the regime of violation to matrix factorisation and verified proposals, treats the equity of curtailment among PV customers as a formal objective, and closes with the ranking of live violations for the operators of the network. Each chapter is accompanied by a notebook that computes every figure and table it contains.