Long-Term Electricity Price Forecast Using Machine Learning Techniques

Abstract

Predicting the performance of energy commodities has long been a global priority for researchers and investors in the Energy sector. Large green field and brown field projects (often exceeding 1bn USD) are financed with locked in capital from the start, and typically take decades to return. Despite being one of the most important aspects of investment decision making, the prediction methodologies used widely today are not sophisticated enough to provide accurate insights for the investors. The new approach was proposed in this research to provide data analytics backed analysis for the performance of energy related commodities using innovative feature discovery methods and machine learning tools. In the presented research, a machine learning model was trained to predict the average monthly price of electricity in the next 5 years with focus on the California State energy market. Data points from 2001 to 2017 were collected and 78 data points are considered for analyses to select the highly-correlated features which could potentially affect the electricity price in the medium to long term. An economic case study is undertaken to understand the correlation of the features, and to avoid multicollinearity. In the next step, the selected features are applied into an S-ARIMA time series prediction algorithm. In addition, several feature-based machine learning algorithms are applied to the data and the results analysed and compared to find the effective forcasting approach. The findings demonstrated promising results for three years future price prediction horizon. Further studies are required to get more accurate electricity results beyond three years horizon.