Allocate kids to schools in base year

[jipclaassens]

• primary school -> age group ?
• secondary school -> age group ?

@cjacobscrisioni geeft nog even zijn code die hij al had gebruikt in een ander project

[cjacobscrisioni]

Demand definitions:

unit<uint32> Facility_definitions: NrOfRows = 4 {
		attribute<string> LabelText: ['Primary_school', 'Secondary_School', 'Elderly_Care', 'Hospitals'], DialogType = "LabelText";
		attribute<string> PopDef: 	 ['y5t9 * 0.8f + y10t14 * 0.4f', 'y10t14 * 0.6f + y15t19 * 0.6f', 'y79t84 * 0.8f + Over85', 'total'];
}

So based on 5y age cohort data:
Primary schools, 4 years from the 5 to 9 group, 2 years from the 10 to 14 group
Secondary schools: 3 years from the 10 to 14 group and 3 years from the 15 to 19 group

[cjacobscrisioni]

For inspiration - this is the code I used to allocate kids to schools and then balance users:

All operations occur within the context of an od matrix (i=inhabited grid cells, j=grid cells marked as a school)

Matrix definition:

unit<uint64> ODMatrix:=
		dijkstra_m64('bidirectional(link_flag);startPoint(Node_rel);endPoint(Node_rel);limit(OrgZone_max_mass,DstZone_mass);alternative(link_imp):alt_imp;od:OrgZone_rel,DstZone_rel,impedance'
				, Network/LengthKM, Network/F1, Network/F2, Network/TwoDirections, jz/NodeRef, Locations/NodeRef, 5f, float32(1), Network/Minutes)

Criteria used to allocate kids:
Locations: subset from 1km grid cells that are flagged to have a location
the distance decay function is ~ dij^b1, b1 = -2 (see notes on transformation of dij)
School dizes depend on observed UK school sizes and neighbourhood student population. Operationalized here as SEM_Def=~ exp(-1.138f + (0.233f*MakeDefined(log(ATT1_j), 0f))) for primary schools

Step 1
A relatively straightforward origin-constrained spatial interaction model based on network distances. HOWEVER the distances in dij are transformed so that only the additional distance for farther schools is taken into account. This is because with the adopted distance decay function, an additional km weighs less in the case of 20 and 21 than in the case of 1 and 2, while we want to minimize distances travelled to some degree.

In pSize1j we establish likely students per school at the location given empirical average school sizes. We consider this an approximation of optimal school size.
In FACj we establish number of facilities at the location (one grid cell can have multiple)
Size1_j is the first guess at school size

attribute<float32>   	dij:	expr = "max_elem(impedance, MinDist)";
attribute<float32>	 	tij:	expr = "max_elem(alt_imp, MinTime)";

attribute<float32>		cls_i (jz): 	expr = "min(dij, i)";
attribute<float32> 		Ai (jz):		expr = "sum(Dj_1 * cij, i)";
					
attribute<float32>		cij:	expr = "max_elem(dij - cls_i[i],1f)^b1";
attribute<float32> 		Oi:		expr = "jz/RelevantPop[i]";
attribute<float32> 		Dj_1:	expr = "const(1,Locations,float32)[j]";
attribute<float32> 		Ai_1:	expr = "float32(1) / sum(Dj_1 * cij, i)[i]";
attribute<float32> 		fij_1:	expr = "Oi * Ai_1 * Dj_1 * cij";
					
attribute<float32>		ATT1_j 	(Locations):	expr = "sum(fij_1, j)";
attribute<float32>		pSIZE1j	(Locations):	expr = "= SEM_def";
attribute<int32> 		FACj 	(Locations):	expr = "max_elem(round(pSIZE1j), 1[int32])";
attribute<float32>		Size1_j	(Locations):     expr = "(ATT1_j / float32(FACj))";

Step 2
This step rebalances student populations given the expected school sizes in Size1. Schools that are larger than the expected school size push students out, schools that are smaller take students in.

attribute<float32>		Dj_2:	expr = "max(MakeDefined(ATT1_j[j] / pSIZE1j[j], 1f), i)[i] / Size1_j[j]";
attribute<float32> 		Ai_2:	expr = "float32(1) / sum(Dj_2 * cij, i)[i]";
attribute<float32> 		fij_2:	expr = "Oi * Ai_2 * Dj_2 * cij";
attribute<float32>		rfij_2: expr = "fij_2 / Oi";
					
attribute<float32>		ATT2_j (Locations):		expr = "sum(fij_2, j)";
attribute<float32>		Size2_j (Locations): 	expr = "(ATT2_j / float32(FACj))";

Size2_j is then the final rebalanced school size

[cjacobscrisioni]

The approach is a combination of empirical observations and best guesses. All published here: https://link.springer.com/article/10.1007/s10109-023-00425-w

When allocating kids to observed schools I would for sure discard the FACj passage, as we actually know the number of facilities per location.

[cjacobscrisioni]

Schools data:

ESPON profecy data 2021 (taken mostly from OSM, separated in primary and secondary schools)

espon_profecy_schools_OSM_2021.zip

ESTAT schools data 2023 (provided by member states, unfortunately not yet separated in primary and secondary schools)

schools_EU.zip

My advice would be to start with the ESPON data while we figure out an easy way to discern primary and secondary schools. I might abuse ChatGPT for the purpose of classifying schools as primary and/or secondary.

[cjacobscrisioni]

Update: I think I can get the EC GPT application to categorize all ESTAT schools as primary / secondary / others. Some work attached as the application prefers data per country. Will proceed to do so in the coming days.

[jipclaassens]

Methodiek geimplementeerd.

To do:

• update kids cohort ipv population Add population groups to model #15
• update school types ipv alle scholen tegelijk Add school dataset to model #14
• update school size (capacity) met actuele waarden (ipv UK kental) Add school dataset to model #14
• controleer alle parameter instellingen