Using Clustering To Reduce Models Required For Medium Term Load Forecasting

Distribution companies sustain their business by providing electricity to its customers reliably. An essential step in this process is forecasting of electricity demand so that the company can develop their business plan for the future accordingly. Distribution companies use a bottom-up approach for medium-term load forecasting which requires several regression models to be developed. This requires a lot of effort to design models for each power feeder. To solve this problem we propose the use of k-means clustering to reduce the number of models required to create a trade-off between effort and satisfactory forecasting accuracy. By using k-means clustering, we group energy consumption for each category and feeder according to their energy consumption pattern. The centroids of each group are then fed to multivariate linear regression (MLR) models. The brute force way is to create a model for each feeder separately. For evaluation, we compare our results with the brute force way. The results show that for about 300 feeders we just require 100 models for a minimal increase of only 5% in average mean average precision error (MAPE) for domestic category.