Load Forecasting Model Using LSTM for Indian State Load Dispatch Centre

Abstract

This paper presents an approach to address the critical challenge of load forecasting in the Indian state of Odisha. Motivated by the necessity for accurate predictions to support efficient planning and operation of the power system network, the work focuses on developing a reliable load forecasting model according to the unique characteristics of Odisha's electricity consumption patterns and environmental influences. To handle this problem, a Long Short-Term Memory based model is proposed with the ability to capture long-term dependencies and handle non-linear dynamics in time-series data. Historical load datasets are collected from the Odisha State Load Dispatch Centre, and meteorological datasets are collected from the National Aeronautics and Space Administration site. The work aims to accurately forecast power demand at 15-minute intervals for both short-term (next week) and medium-term (next month) horizons. Through comparative analysis with traditional methods such as Gaussian Process Regression and Artificial Neural Network, the proposed approach demonstrates superior performance in terms of accuracy and reliability. One year of the dataset (from January to December 2022) is considered as a training dataset to forecast next year's January 2023 load demand at every 15-minute intervals. The LSTM model achieves an absolute error range of ±10 MW during testing, with a mean absolute error of 5.9443 MW and a Mean Absolute Percentage Error of 0.2134%, outperforming the existing models. This research contributes to advancing the reliability and efficiency of power system operations, offering valuable insights for optimizing load forecasting strategies in Odisha and similar regions.