Modified LSTM with Memory Layer for Power Grid Signal Classification

In processing of signals coming from power grid system, the major purpose is to filter out the key signals that are highly related to power grid fault or breakdown. Power signal analysis and incident summary are normally done by people. The obvious issue with manpower is when multiple incidents take place and huge number of signals emerge, processing all signals manually in a short time becomes virtually impossible. Attempts to handle similar massive information situation with neural network models have been proven useful. However, ordinary neural network models that learn and process signals usually suffer “memory loss” as the sequence gets longer, which leads to inaccurate classification. This paper proposes a modified version of long-short-term-memory (LSTM) network that can mitigate the memory dependency issue of recurrent neural network(RNN) and basic LSTM structure. The focus of the thesis is how the modified LSTM network can utilize the additional layer of memory slots to store the entire output sequence from the last LSTM cell instead of mere output value in basic LSTM architecture. The modified model can incorporate all previous cell information and prevent memory loss in the modeling process. To substantiate the merit of the LSTM model with stacked memory layer, a signal classification experiment is done with both modified and basic LSTM models and the results are shown.