Use High Frequency Data for Smart House and Microgrid Optimization

Great news! UC Berkeley students are making predictions about future energy usage for home appliances. By serving more efficient and smarter energy, they hope to make people’s lives better and more sustainable. They obtained their data from ETH Zurich websites, which includes data points recorded every second by smart meters installed in home appliances with a time span of 8 months. Before diving into modeling, they first cleaned and aggregated the data to create one master pandas dataframe for each house from 3 csv files containing 20.7 million rows of data in total. And then, they did exploratory data analysis and studied seasonal decomposition. They explored consumption patterns and behaviors through times series visualization in particular using Plotly and Cufflinks. After learning more about energy consumption patterns through visualization, they decided to make predictions of energy consumption of home appliances by applying Arima & Sarima model, Ridge regression and Tensorflow(deep neural network). By assessing the performance metrics, they found out that SARIMA was the best performance model among all their models. Ridge regression was able to somewhat accurately capture the consumption pattern, and Tensorflow fails to capture the time series patterns, which are lost in randomizing test and training data. They also learned that it is preferable to use DNN to predict the consumption range as categorical (Classification) instead of continuous response variable.

By learning people’s energy consumption patterns, companies can use their model to adjust the usage of electrical appliances based on model predictions. For big companies who are on budget, they can easily adjust their usage of electricity in the office. For example, coffee machines can be turned on right before employees arrive their offices in mornings. On the other hand, electricity companies can use people’s energy consumption patterns to better distribute energy through microgrid. In energy consumption peaks, they can supply more, and vice versa. Additionally, abnormal energy usage can also be detected when there is a discrepancy between our prediction and the actual consumption. The energy distributor can turn off energy switches automatically for customers to prevent accidents. They are so excited about all the potential benefits that this project brings and are looking forward to transferring all the work to the next semester for future user interface developing.