Machine learning for energy load forecasting

With an increasing penetration of renewables into energy markets, it is desirable to have a flexible grid in order to match large fluctuations in supply to a volatile power output typical of renewable supply. Hence, it is imperative to accurately forecast power load demand. The recent emergence of big data analytics and machine learning techniques have shown great success in a wide range of regression problems in varied industries and various data can be harnessed by the energy industry to better understand likely energy loads placed upon the system. This paper presents a comparison of several regression models which can be used for accurate predictions of energy load given environmental feature data. Here we show that dynamic Gaussian Processes can be used as a powerful tool taking into account the non-stationarity of the data under analysis. This regression model was compared Neural Networks, used most extensively in the industry, and linear regression models to give an idea of their comparable accuracy. However, it was noted that the dynamic Gaussian Process were inferior to a Neural Network when training for huge datasets due to their high relative computational cost, increased uncertainty with projection time, and large memory usage. Though primarily used for dynamics problems, there are a range of non-stationary problems which could benefit from the use of a dynamic Gaussian Process of which this paper just presents one. It also considers online learning models be used for real time forecasting.