Predicting the Power Consumption and Operating Rate of Enterprises in Major Public Events

The power consumption of enterprises in major public events may significantly different from non-event days, which brings new challenges for the existing power distribution system. Conventional approaches to predicting power consumption are mainly based on statistical learning methods, such as a particular fitted distribution, logistic regression, decision trees, etc. However, the customer's power behaviors change significantly during the major public events, which may lead to suboptimal performance for existing methods. To overcome these challenges, we propose a novel long and short-term memory-based attentional algorithm to accurately predict the power consumption and corresponding operating rate of enterprises in major public events. In particular, we firstly employ long term memory gate to learn the most important historical pattern and forget the irrelevant behaviors. Then, the short-term memory is leveraged to increase the importance of recent patterns. Lastly, compared with the conventional method that gives equal weights to different slices, we design an attentional prediction network to dynamically adjust the weights of long and short-term patterns. We optimize the proposed end-to-end deep learning model by standard stochastic gradient descent (SGD) algorithms.