Abstract
With rapid advancements in renewable energy sources, billing mechanism (AMI), and latest communication technologies, the traditional control networks are evolving towards wise, versatile and collaborative Smart Grids (SG). The short term power load forecasting of individual as well as group of similar energy customers is critical for effective operation and management of SG. Forecasting power load of individual as well as group of similar energy customers is challenging compared to aggregate load forecasting of a residential community. The main reason is the high volatility and uncertainty involved for the case of individual and group of similar energy customers. Several machine/deep learning models have been developed in the recent past for forecasting load of individual energy customers, but such explorations are ineffective due to the requirement of one trained model for every energy customer, which is practically not feasible. We plan to build a deep learning model using convolutional neural network (CNN) layers in pyramidal architecture for effective load forecasting for a group of similar energy-profile customers. Initially, we grouped a subset of energy customers from database of Smart Grid Smart City (SGSC) into clusters using DBSCAN approach. The CNN layers are used for extracting feature from historical load of each cluster. The extracted feature of similar energy-profile customers (grouped based on clustering) is combined to make training-databases for each cluster. We have used the power load data from SGSC project, which contain thousands of individual household energy customers data. The developed Pyramid-CNN model is trained based on these sets of databases. The trained model is evaluated on randomly selected customers from few clusters. We obtained significantly improved forecasting results for randomly selected user from different clusters. Our adapted strategy of clustering based model training resulted in upto 10 percent MAPE improvement for the energy customers. The essence of our work is that energy customers can be grouped into clusters and then representative model could be developed/trained, which can accurately forecast power load for individual energy-customer. This approach is highly feasible, as we do not need to train a model per energy customer and still achieve competitive forecasting results.