Day-Ahead Electricity Market Forecasting Using Ensemble Machine Learning Approach

Abstract

Abstract: Due to the growing complexity and volatility of the energy market, there is a need for advanced forecasting methodsto assist operational decision-making and risk management. This study offers a complete machine learning framework forpredicting key energy market variables using time series high-frequency data of Day Ahead Market (DAM) in India. Weanalyzed market data spanning the whole year of 2024. This market data has 35,040 observations at 15-minutes frequency. Wefocus on three variables: market clearing volume weighted (MCV), scheduled volume and market clearing price weighted(MCP). Our methodology uses several advanced techniques to generate 32 new variables based on temporal patterns, lagrelationships, rolling statistics, and market-specific indicators. Numerous models have been developed and evaluated. Thisencompasses four models, which are random forest, linear regression, a moving average baseline, and an ensemble approach.Furthermore, a Long Short-Term Memory (LSTM) neural network was implemented to capture long-term dependencies in thedata. Our system is not only more accurate but also computationally efficient. Model training times were reduced from hoursto less than 10 minutes and general performance remains strong. When predictions for 15 mins are accurate, it is much easierto run things in a way which has easier trading strategy. The research shows that advanced machine learning and featureengineering can provide very good predictions for energy market applications. This study will definitely increase knowledgein energy market analysis and provide necessary information to the market players to use it in improving operations in a tradingenvironment which is becoming increasingly dynamic.