Which method to use for load forecasting I have smart meter data set that has consumption readings collected over a year and a half for every 30 mins. What I am trying to do is short term load forecasting. The data set has just three columns CUSTOMER_ID, date/time-stamp and CONSUMPTION. 
Which method should I use to get a higher accuracy with just the available data? Should I use uni-variate or multi-variate time series (or) SVR or ANN? And please do suggest if any other algorithms can be used.
NOTE: the data set doesn't have any information regarding weather.
 A: I have been working this problem for a year now. ANN gives the best results but if you are not familiar with the area you might have difficulties with the implementation. On the other hand SVR are very easy to implement but you must spend some time finding the optimal parameters (you must do that and for the ANN but it is easier) and the results are a little worse than ANN(it depends but this is the general rule). You can use an algorithm from the ARIMA family but they have the poorest results.  The most important step is to find the best features to use. For example you can use the values of the consumption before n days as features, but you have to play a little with different values of n. One more algorithm that you can use is Gradient Boosting Regression. Also another factor is the type of building that you are doing the forecasting, the weather, the occupancy, but you do not have that data. In the end the only way to find the best model is to try them all. There is a ton of bibliography on the subject.
A: You have 48 readings per day for a year and a half. We have found that by combining memory effects (ARIMA) and deterministic effects like day-of-the-week , level/trend effects with daily effects as a predictor some reasonable results can be developed while adjusting for anomalous readings and daily trends. The trick is to develop a daily total forecast that also uses the aforementioned structure while optimally incorporating lead and lag effects around holidays , particular days-of-the-month , weekend effects and even weeks within a month , etc . Model identification procedures to reconcile 1/2 hour forecasts and daily forecasts and of course the lead/lag holiday effects are available in commercial products like AUTOBOX ( which I have helped to develop ) and to a much lesser degree in products like SAS and SPSS. In my opinion free software within R is insufficient to deal with the model identification opportunities provide by data that you have available but it is free and may provide a start.
I would suggest that you approach serious providers of time series software and deliver your data and ask them for a free demo. Review the proposals and then either make or buy the solution that is best for you.
