i want to apply short term forecasting on the smart meter data
For example: Suppose the data set are
T1 : 10
T2 : 20
T3: 50
T4: 30
T5: 70
T6: 50
T7: 80
T8: ?????
How can i get the consumption value in T8 by neural network and SVM?
I need equation or tool detect the consumption value in T8.
Thanks in advance.
Sounds like you want some form of a recurrence neural network. These can be seen as a time series style of neural network. The R package RSNNS can do these pretty well. If you want to use an SVM then you will want to look into support vector regression : http://www.svms.org/regression/SmSc98.pdf . Although if your real data only consists of 8 time points estimation is going to be pretty difficult.
Thanks for your help.
Are there other methods for applying forecasting on small data set ??
How many values (data set size) required to NN and SVM to give a good results??
For a data set this small I would probably just use a MA (Moving Average) model. Predictions with small sample size data sets are VERY sensitive to the amount of variation in that data set. For ANNs you want a fairly hefty data set with quite a few time points. I am hesitant to say an exact amount but I would want at the very least 50. With the SVM maybe less but these numbers are really pushing it when it comes to estimation.
You mentioned that the data is from smart meters - then you probably have date/time associated with each reading. The smart meter may be for a building that has some periodic behaviour and the consumption may be related to time of the day, season..., so you may want to include those. Moreover, you may want to include weather variables, such as temperature and humidity.
Example suing SVR and NN: http://www.sciencedirect.com/science/article/pii/S0378778815304461
Example with M5 model trees with weather attributes: http://publish.uwo.ca/~kgroling/papers/Demand_Forecasting_M5.pdf
With smart meter data, the data sets are typically not small (there ate some available online); however, if you have truly small set SVR and NN will not work.
ARIMA is another approach often used with time series
- Badges
- Science topic