When calculating custom energy savings for a consumer, inputting a detailed consumption history can become a problem. Instead, we can put the user into a class based on more easily obtainable broader information and use pre-set data for that class. We tackled the issue of creating classes by clustering different consumers based on their yearly power consumption timeseries using KMeans. Then we tested how accurately we can sort a consumer into its correct class based on different parameters (or features) using different methods. We found the best combination of features to be the quarterly maximum power consumption and the total energy consumption of each month. The best single feature was the total energy consumption of each month. The best method turned out to be Random Forest. This allows the user to input less data while at the cost of lower accuracy.
|