Window Function Lag/Lead supported on Filter (Not equal) for FloatField

[Gabriel Oliveira]

I'd like to be able to reproduce the following example in Django, but on the Temperature column
(​https://fle.github.io/detect-value-changes-between-successive-lines-with-postgresql.html)

Given a table

db=> SELECT * FROM weather ORDER BY day DESC;

day	temperature	rainy
2014-04-08	20.0	f
2014-04-07	20.0	f
2014-04-06	16.0	t
2014-04-05	16.0	t
2014-04-04	16.0	t
2014-04-03	22.0	f
2014-04-02	22.0	f
2014-04-01	22.0	t

I'd like to show dates only if the temperature changed:

day	temperature
2014-04-08	20.0
2014-04-06	16.0
2014-04-03	22.0

On pure PostgreSQL, this translates to:

SELECT
    w1.day, w1.temperature
FROM
    (SELECT
        w2.day,
        w2.temperature,
        lead(w2.temperature) OVER (ORDER BY w2.day DESC) as prev_temp
     FROM
        weather w2
     ORDER BY
        w2.day DESC) as w1
WHERE
    w1.temp IS DISTINCT FROM w1.prev_temp
ORDER BY
    w1.day DESC;

I could accomplish the inner query by using the new Window Functions:

Weather.objects.annotate(prev_temp=Window(expression=Lead('temperature'), order_by=F('day').desc()))

Now my problem is use this annotation to filter only when temperature differs from prev_temp
(in order to accomplish something similar to the "temperature IS DISTINCT FROM prev_temp")

When I try to use the available filters, the following errors occurs:

Weather.objects.annotate(prev_temp=Window(expression=Lead('temperature'), order_by=F('day').desc())).order_by('-day').filter(temperature__ne=F('prev_temp'))

Gives the error:

FieldError: Unsupported lookup 'ne' for FloatField or join on the field not permitted.

Another try:

Weather.objects.annotate(prev_temp=Window(expression=Lead('temperature'), order_by=F('day').desc())).order_by('-day').filter(~Q(temperature=F('prev_temp'))))

Gives the error:

ProgrammingError: window functions are not allowed in WHERE

It's in fact a PostgreSQL error, because the generated SQL query tries to pass the LAG function inside the where clause.

How may I accomplish that, even if I have to use the extra fields, or even RawSQL class ?