Detecting anomalies in wholesale electricity day-ahead market bidding data using LSTM-network

Abstract

In this thesis it is studied how neural networks can be used for anomaly detection in day-ahead wholesale electricity market trading data. Economics of electricity markets lays a foundation on detecting distinctive patterns in supply behavior reflecting market manipulation, such as economic and physical withholding of production capacity. The impact of market abusive supply behavior is studied on shapes of supply curves. A neural network model is used to provide score to measure stationarity of bidding behavior. An unsupervised machine learning framework for anomaly detection is set up by using a rolling window approach. 24 hours of high dimensional supply trading data is used as input to make prediction one hour ahead. Prediction errors of every individual hour are used as a time series of anomaly score, which is thoroughly analyzed in the light of signs of market manipulation based to the literature of electricity market economics. The study is conducted on two years of anonymous aggregated day-ahead trading data collected by The European Union Agency for Cooperation of Energy Regulators received from Energy Authority of Finland. Idea is to fit neural network to the data to estimate how supply curve of an hour would look like conditional on 24 previous hours and external variables. Neural networks are used for the estimation as they are capable of modelling non-linear spatial dependencies in the data. LSTM model is further chosen because it is designed to handle long term dependencies in the data. If the prediction errors are low enough on average, it can be assumed that the model can capture stationary behavior in the data and outliers can be assumed to result from changes in data generation process. If model can predict supply curves well enough on average, large prediction errors can indicate that something unexpected has happened in the markets. LSTM-model is trained to make rolling window predictions using 5-fold walk forward validation approach, where chronological order of the data is maintained to mimic real life prediction scenario. Early stopping is used to prevent overfitting. Hyperparameters are chosen via grid search likewise using 5-fold walk forward validation. Two major distinctive types of supply behavior are identified from the literature, economic withholding and physical withholding. Their impact is studied on supply curves and is paid attention in the analysis of anomaly score. Mean absolute error of individual hour is chosen for anomaly score, which is referred as h-MAE. Performance of the model is compared to one used by Guo et al. (2021) in similar function of predicting supply curves. Method is promising in detecting out-of-ordinary supply curves, based on thorough statistical survey of the results and brief qualitative survey, in which a confirmed market violation was detected, as well as erroneous period in the data. Linking market manipulation to the anomaly score directly proves difficult. However, the method offers a noteworthy possibility to surveil and study supply curves in day-ahead market as it benefits from enormous amount of high-dimensional data and is capable to take account the spatial and temporal non-linear relations of the supply curves. This Master’s thesis was done in affiliation with Energy Authority of Finland.