Natural gas consumption forecasting

Hi, I have 4000 hours of the natural gas consumption in one station. This information is the hourly consumption profile for every day during the last six months. I have information about if the day is holiday or not, the weekday and the hour of the day (1-24). Given the first 12 hour of one day, I would like to forecast the other 12 hours. How can I do this ?

My first idea, was to compare the 12 hours or part of it with all the records I have, take the top three most similar, and use their weighted profile (in function of the similarity) to forecast the other 12 hours.

for i = 1:NDias
    PerfilDia_i = DiaHorario(i).data(HorasSel,VolC);
    Perfil_Norm_i = (PerfilDia_i - min(PerfilDia_i))./(max(PerfilDia_i) - min(PerfilDia_i));
    DistPerfiles(i) = norm(Perfil_Norm_i - PerfilNorm0);
end

[ DistOrdered, I ] = sort( DistPerfiles );
DistOrdered(1) = []; I(1) = [];
NPerfiles = 3;
Pesos = DistOrdered( 1:NPerfiles );
Pesos = 1 - Pesos / sum( Pesos );
Pesos = Pesos / sum( Pesos );

PerfilTotal = 0;
for i = 1:NPerfiles
    Perfil_i = DiaHorario( I(i) ).data( :, VolC );
    Perfil_i = (Perfil_i-min(Perfil_i(HorasSel))) / ...
        (max(Perfil_i(HorasSel))-min(Perfil_i(HorasSel)));
    PerfilTotal = PerfilTotal + Perfil_i * Pesos(i);
end

ConsumoReal = DiaHorario(DiaTest).data(:,VolC);

ForeCast = PerfilTotal * (max(PerfilDia0) - min(PerfilDia0) ) + min(PerfilDia0);

This idea has given to me good results, but I would like to use something more complex and surely, accurate. I have read that MATLAB has powerfull tools for this. What do you recommend me to do ?

Roberto E.

[dpb]

The obvious other variables to include would be some of those relating to demand -- like temperature is one major driver if this is a generation plant you're talking about.

[Roberto]

No, I am talking about a distribution system. This is one of the measure and regulation stations.

[José-Luis]

There are many many ways to approach this problem. You could develop a model, you could do a regression, you could do a bayesian approach, etc... What's the best way to go? Hard to say. If you're not interested at all in understanding how the system works you could even try artificial neural networks.

[dpb]

...talking about a distribution system.

Still only distributes what is demanded so the demand drivers are still going to be highly correlated and predictive...

Answer 1

[Yifeng Tang]

I feel like ARIMA (https://en.wikipedia.org/wiki/Autoregressive_integrated_moving_average) may be a suitable tool/model for this case. What I think may be a risk though, is whether natural gas consumption has a yearly seasonality to it, e.g. colder months see more consumption. If so, 6 months of data may not be enough, or you are likely to get the "average" part wrong.

Below are some resources I've found on MathWorks Help:

https://www.mathworks.com/help/econ/specify-arimap-d-q-models.html

https://www.mathworks.com/support/search.html/videos/creating-arima-models-using-econometric-modeler-app-1520280167404.html