Deep Learning-based Unsupervised Anomaly Detection on Energy Consumption Data

Time series anomaly detection is the process of identifying deviations from expected patterns within sequential data points over time. It is a fundamental task in all those scenarios where analysing and understanding temporal trends is essential. This field leverages advanced statistical and machine learning techniques to detect irregularities, spikes or unusual patterns in time series data. Nowadays, it has become of great importance to perform anomaly detection on energy consumptions, in order to identify and prevent a wasteful use of energy for a better management of LECs in EU. This thesis investigates state-of-the-art Deep Learning approaches applied on energy consumption data in a completely Unsupervised manner. The dataset used for the development of the chosen models is taken from the Large-scale Energy Anomaly Detection competition hosted by Kaggle. The methods explored work as to reconstruct or forecast the normal behaviour of the input time series; then, anomaly detection is performed by classifying as anomalies all those output data points which differ from the expected values more than a certain threshold. The goal of the thesis is to understand the applicability of these methods in absence of annotated data. In many real-life scenarios, in fact, datasets are not provided with indications of whether the contained observations are to be considered as anomalies or not. This makes the anomaly detection task harder. Without labels, in fact, one has no way of identifying and removing anomalous data points from the dataset which is going to be used to train the models: they will thus learn to recognize abnormal patterns in data as normal energy consumption behaviour. This work explores various Deep Learning techniques and thresholding methods in order to enhance the development of more resilient and effective models for practical applications in energy management.