Feature Identification Approach for Big Appliance Data Based on LSTM Neural Network

• Main Authors: CHEN, YAN-RONG, 陳彥融
• Other Authors: LAI, CHIN-FENG
• Format: Others
• Language: zh-TW
• Published: 2017
• Online Access: http://ndltd.ncl.edu.tw/handle/6p3wq4

碩士 === 國立中正大學 === 資訊工程研究所 === 105 === Energy saving and carbon dioxide reducing are significant issue now. We can use the smart meters to monitor the power usage of applications to pay some offer to environment protection. Although the electrical feature collection can be done by smart meters, applications need to be labeled by human. There would be some mistakes as manually labeling or non-unified names. To solve this issue, we propose a novel approach to conquer this problem automatically. In this paper, we use the LSTM neural network in our model to achieve application recognition. First, we label part of data manually as input data to train our model to find the relationship between the name and the electric feature. Then we use the unlabeled data as input to let the model to recognize them. Now we have GPGPUs to help us to accelerate the parallel computing and the training progress of neural networks. Through the parameter adjustment policy, we set our model with 512 LSTM nodes per hidden layer, 12 hidden layers, no dropout, ReLU as activation function, learning rate to 0.000001. The result shows that we achieve 88% accuracy in random data testing and 83.6% accuracy in sequential data testing of single application.