Abstract
An essential benefit of demand response is the avoided necessity to build new power plants to serve increased demand that occurs a few times during a year or to mitigate the energy imbalance caused by the volatile renewable energy. Customers play a pivotal role in the demand response program, and their behavior illustrates a highly uncertain pattern. To quantify the uncertainty, we follow the approach of data-driven probability distribution modeling by training a mixture density recurrent neural network, which outputs the probability distribution of the demand reduction. The parameters of the obtained mixture distributions are time-varying, thus capturing the temporal impacts of customer behavior on the customer load reduction. Using specific statistical metrics, we compare the performance, i.e., the quality of scenarios generated, of the Gaussian mixture distributions with that of the single Gaussian or ormixture distributions obtained by fitting the raw consumption reduction data. The proposed methodology is then applied to an optimal customer selection problem, which is formulated as a risk-averse stochastic knapsack problem. The results indicate that the generated mixture distributions are better suited for quantifying a customer’s consumption reduction and accurately encapsulate the underlying spatiotemporal trends in the customers’ reduction pattern.
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have