Medium and Short Term Energy Forecasting using LSTM Neural Network Method for Gujarat State

A stable power system requires balance between generation and demand. The power system under World Bank initiative has restructured in majority of countries throughout the world. Implementation of Availability based tariff (ABT) and establishment of Energy exchange, congestion in network etc. has made it necessary to forecast the load condition accurately. Load forecasting is required to be performed for different of timespan, i.e. short, medium and long term. Some machine learning based approaches were used in recent past to address this problem, however taking in account of the dependencies of load demands on load density, city and industrial plants, local and global policies there is scope of forecasting precision. This paper proposes a deep learning based neural network architecture, i.e., Long-short memory term (LSTM) neural network for energy forecasting in short and medium term with the collected dataset from State Load Dispatch Centre (SLDC), Amreli, Gujarat. The dataset includes three months (May-July 2014) of energy consumption at hourly basis. Short-term as for the next 24 hours and medium-term forecasting for over 9 days are performed with root mean squared error 88.05 and 87.97 MW, respectively. Furthermore, forecasting is also performed for holidays, preholiday and post-holiday periods which exhibited even less RMSE.