I’m working on an model for auto dimensioning district heating pipes for new district heating areas (new customers). I have energy consumption data on hourly basis and describe data about these consumers (e.g. building year, renovation year and floor area) for a lot of district heating consumers. I want to predict dimensional load (energy consumption) for new building connected to the grid based on the describe data (e.g. building year, renovation year, floor area). I have through about using supervised or unsupervised learning, but do not know which of them that fit my needs? Is there an mode (e.g. clustering, SMV, …) that are better than other for this purpose?
Building on The Lyrist answer, I would try to organize all that data and create as many instances as possible of Input-Output. If I understand your case correctly, the input would be the multiple data point that you feel have an impact on the load (if you're not sure about it's impact, chuck it in for now) that you would set up as a vector of values. Then the output would be a single value of load, such as Kwh/day (that doesn't sound right), but just get one 'value' you're interested in predicting. Don't forget to keep training, validation and test sets.
Now you can frame your problem both as a classification problem by creating categories, such as from 0 to 10, 11 to 20, etc, or go for a regression problem, predicting a continuous value. Do lookup past application of similar problem, as they are, to some extent, application specific. Note that the some algorithm might work better with regression, other with classification. Personally, I think that creating buckets of classification might be easier, unless predicting an exact value is important to you.
Now regarding the specific model to go for, it usually is recommended to start with simple model (SVM's is a popular one) and build to more complicated model (kerneled SVM), if you enough data! Resist the temptation of going straight for a 12 layers MLP, as those old SVM's.