Using LSTM neural network for power consumption forecasting

Abstract

Massive integration of distributed renewable energy sources (DRES) into the power grid will eventually change the supply behavior of the traditional power system. The RES output is obviously unstable, so the system’s reliability should be considered carefully. This is a process of accurately balancing generation capacity to the demand of the consumers. Storing generated energy is a huge cost, so energy is lost in the transmission networks during off-peak times, in contrast, the system suffers from a deficiency of energy during peak times which leads to the disconnection of certain areas from the network. This situation is a main source of damage to the power system and economic losses for the utility. This work analyzes power consumption data of the Andijan region of Uzbekistan on a daily frequency. Different lengths for input sequence data to the network data were selected according to the autocorrelation of the data. The results showed that longer sequence data is beneficial to the LSTM network in case of strong autocorrelation.