Tidy Time Series Anomaly Detection for Load Forecasting


Accurate load forecasting is vital for effective energy management, as forecast underestimation can result in blackouts, whereas overestimation may result in energy wastage. However, the quality of historical data for load forecasting can be affected in several ways, such as data integrity attacks, missing values, incorrect readings, technical aberrations. These issues make data unreliable and untrustworthy and can have a direct impact on forecast accuracy and subsequent decision making. This work develops a framework for detecting anomalies in tidy time series data. An anomaly is defined as an observation that is predicted as very unlikely given the robust time series forecast models. The algorithm works with tidy temporal data provided by the tsibble package and produces an outstable, a tsibble with flagged anomalies and their degree of outlierness. An approach based on extreme value theory is applied to residual series in order to calculate a data-driven anomalous threshold. The proposed framework can also provide a cleansed tsibble that closely integrates with the tidy forecasting workflow used in the fable package. A number of different approaches are available for the data cleansing process. The wide applicability and usefulness of this proposed framework in load forecasting will be demonstrated using various synthetic, real-world, and publicly available benchmark datasets including data from Global Energy Forecasting Competitions. This framework is implemented in the open-source R package outstable.

Jun 30, 2021 12:00 AM — 12:00 AM
Virtual ISF 2021
Virtual Conference
Priyanga Dilini Talagala
PhD in Statistics

My research interests include Computational Statistics, Anomaly Detection, Time Series Analysis and Machine Learning.