Power Consumption Predicting and Anomaly Detection Based on Long Short-Term Memory Neural Network

Identification of abnormal power consumption data is a critical and difficult task to guarantee efficiency and reliability of the smart grid. This paper proposes a power consumption predicting and anomaly detection approach based on LSTM (Long Short-Term Memory) neural network. Compared to the ARIMA (Autoregressive Integrated Moving Average) algorithm, LSTM model demonstrates a decrease of 22% in forecasting error. Our simulations are performed based on actual power consumption data of ten thousands of users from Chongqing, China in the period of one year. Different from previous studies, we predict electricity consumption and detect anomaly in long-term prospective. We focus on seasonal and monthly trends, instead of power consumption pattern from typical commercial and residential buildings at metering granularity at once every 15 or 30 minutes. Electricity theft identification is significantly improved compared to previous unsupervised algorithms.