Adapters are specific to the publisher. Everything above the schema is not, and that is the whole point of the package. This page describes two analyses that become possible only after harmonization.
Literature reviews on siting network infrastructure list available connection capacity as a criterion that is frequent and at the same time unmeasurable, because of the absence of the data in open sources. It is measurable, provided the publisher discloses both the available capacity at the node and the composition of the substations belonging to the node. In the panel, one publisher does so.
from gridqueue import station_node_map, assign_to_nodes, node_loading
mapping = station_node_map(panel) # station -> group of stations
panel = assign_to_nodes(panel, mapping) # assignment of applications to groups
L = node_loading(panel, capacity_field="capacity_injected")
L[["node_name", "applications", "capacity_requested", "capacity_available", "loading"]]
The key is the group of stations, not the node row. The publisher lists the same group twice, separately for the generating and the consuming direction, so a mapping based on the row identifier would make every substation ambiguous. That would be an artefact of the document layout, not a property of the network.
The assignment is by name, not topological. It rests on the agreement of the names disclosed by the publisher, not on a model of the network, and its correctness was not validated against an external reference set, because none exists. The matching status is returned explicitly and distinguishes five cases:
| status | meaning |
|---|---|
EXACT |
substation name equal to the location after normalization |
EXTENDED |
the location contains the substation name as a contiguous subsequence of tokens |
LINE |
an object planned on a line between substations, not at a node |
AMBIGUOUS |
more than one group matches |
NONE |
no hit or no location |
The LINE category exists because entries of the kind “planowany RS w linii 110 kV
relacji A, B” describe an object lying between nodes. Matching by name
would assign their full capacity to whichever end of the line happens to appear in the
list of substations, which is arbitrary. Such applications are left unassigned and
counted separately.
A node with zero disclosed available capacity gets a quotient of NaN, not
infinity: the quotient there is undefined, and averaging it would be an
error. The applied capacity is nevertheless retained for it, because a queue to a node
with no free capacity is information, not missing data.
An application with no settlement date is not missing data: it is an observation truncated at the publication date of the edition. Computing the median only on complete cases understates it, and the more strongly the longer the real processing time is, that is most strongly exactly where the result matters to us.
from gridqueue import processing_time, kaplan_meier
d = processing_time(panel) # time_days + event (1/0)
km = kaplan_meier(d["time_days"], d["event"])
km.as_dict()["median"], km.as_dict()["censored_share"]
The Kaplan-Meier estimator was implemented directly, with Greenwood variance, so as not to introduce a dependency on a survival analysis library for the sake of one estimator. The interval for the median is determined by inverting the confidence band of the curve.
linkage_audit measures precision: whether the linked pairs concern the same
case. linkage_recall measures recall: how many pairs the key missed. The latter
are the more dangerous, because a broken pair appears as one “departed” row and one
“new” one, so it looks like an event in the register when it is an artefact of the
linkage.
from gridqueue import linkage_recall
r = linkage_recall(edition_a, edition_b)
r.as_dict()["key_sensitivity"]
r.recovered_keys["different_key_fields"].value_counts()
The identity fields used for recovery have to be disjoint from the key. A field shared with the key will not detect a pair broken by a change in exactly that field, because the identity diverges together with the key. Measured on a pair of editions of the transmission operator: adding the location to the set of identity fields raises the apparent recall from 0.9652 to 0,9757, and in reality hides ten pairs.