A Basic Query on neural network
Greg Heath
>I have seen your many messages on Matlab forum and hence am asking you for help.
>I am doing a project on Electricity price forecasting using neural networks (back propagation, radial basis). I am highly confused on which way i will be giving the input. I have an hourly price data for a whole year (24 X 365 matrix) divided into 4 seasons...one of which is a 24 X 92 matrix....i.e. 24hrs by 92 days . I have to predict the prices of the next 24 hours based on the previous 3 days data. I am using 75 days for training and 17 for testing.
>My query is... should my input matrix be [72 X 75] and output matrix be [24 X 1]. I am getting matrix mismatch error by this. I am not very clear on when we say for e.g no of input neurons is 5 ..does it mean 5 rows or 5 columns when its a matrix ??
>
The sizes of the input and output matrices are explained
in the documentation
doc newff
help newff
size(p) = [72 89]
size(t) = [24 89]
size(ptrn) = [72 Ntrn]
size(ttrm) = [24 Ntrn]
size(ptst) = [72 89-Ntrn]
size(ttst) = [24 89-Ntrn]
Ntrn ~ 60 or lower might work.
Hope this helps.
Greg
baldrick
I recommend you find out a bit more on how the price of electricity is
determined first. Simply having the previous 3 days prices is not
going to give you a great model.
1. Price is determined by demand.
2. Demand depends on:
A) the time of day (due to lighting load which varies seasonally,
office hours which are pretty static, evening peaks when everyone is
cooking)
B) day of week - weekends have lower loads due to not many people
being at work.
D) seasonality - mainly due to lighting
E) the weather (also seasonal - heating/air conditioning load
dependant on temerature)
If this is merely an academic exercise then go ahead, but if you
actually want to build a good model then you need to build a causal
model rather than a time series model.
Hope this helps.