Dear Rajeev,

The work done on load forecasting techniques is very broad and diverse indeed. Furthermore, I doubt anyone can give you a best technique right off the bat as this is highly dependent on various factors. For example, do you wish to do short term (e.g., day ahead) or long term (e.g., expected load profiles for grid planning/expansion) forecasting. What kind of scale are you interested in (household, neighbourhood, city, or country). Do you have an idea what the main influencing factors are (e.g., weather, and type of day) and does it matter that you can relate these to the actual forecast? Do the forecasts need to be point forecasts or are you looking for confidence intervals?

I hope the above makes it clear that there is no easy answer to your question. Unfortunately I cannot currently give a better answer than to review the body of work out there and attempt to find a method which best suits your needs. I have attached a rather recent survey on load forecasting which might be helpful. Note that, as is noted in the conclusion of this survey, a broad and recent comparison between load forecasting methods appears to be lacking. Furthermore, the soft computing techniques (e.g., neural networks, and fuzzy logic) seem to have gained a fair amount of popularity in recent years

http://ieeexplore.ieee.org/xpls/abs_all.jsp?arnumber=6508132&tag=1

Dear Neural Network is the basic load forecasting technique, however researcher are using many hybrid techniques of neural network and evolutionary algorithms these days.

This might be interesting to you:

https://repositorio-aberto.up.pt/bitstream/10216/59109/1/000144800.pdf

In my view Support Vector Machines (SVM) technique is the best

I used SVM and it is more powerful  than any other technique

Dear Monsef Tahir and Balaji,

                                                     In which way you found SVM is better technique. 

Do you have SVM simulation model. 

Time series analysis with weightings allocation was the traditional method used  based on past load data or load duration curves  aggregated over may years. Modern tools especially in Artificial Intelligence (AI), including ANN have  proven to achieve better results. The subject is broad and several contributors to this thread have alluded to that already. My judgment is that you  chose or adopt a   methodology that is skewed towards the type of load characteristics you are looking at: peak demand, base load, generation plant portfolio, spinning reserve capacity.

I have performed demand forecasting using Neural networks and linear/polynomial regression methods individually. I find Polynomial regression is simple, less complicated and effective.

The Holt-Winters method is a relatively simple but effective algorithm -

http://users.ox.ac.uk/~mast0315/TripleSeasonalEJOR.pdf

The fundamental time series analysis method remains useful for most power system operators, with weightings allocation to different indices, special events, economic growth forecast GDP and so forth. Several complex methods and algorithms have been proposed in recent years. I recommend you look at the original methods used before high-power digital computing became the norm, and interrogate any proposed methods based on the fundamental theory of load forecasting. The level of modernization of the grid in question does matter. With the availability of smart meter data in several cases, load modelling has become much easier using these modern tools with accuracy.