Demand Response (DR) programs are designed to reduce energy consumption for relatively short time periods (e.g., a few hours per event). It has been widely recognized that DR can help to meet both reliability and market needs. In order for DR programs to achieve their full benefits, however, it is critical for utilities to accurately predict the reduction in energy consumption during events and increases due to the rebound effect after events. Currently, DR prediction is performed based on the historical energy consumption data without the impacts of anomalous data points. However, days with anomalous energy consumption, such as when the consumer is on vacation, can bias analysis of historical consumption behavior, and therefore significantly decrease the accuracy of DR prediction. This is especially the case when anomalous days occur during DR event periods or baseline measurement periods, where there is a small sample size for evaluation. This paper presents a method to accurately identify anomalous days for individual premises so that they can be removed from the premise data. This will enable more accurate assessments of energy consumption patterns, including normal usage, consumption baselines used for billing, and DR estimation algorithms. Several different methodologies for anomaly detection are discussed. These methods either utilize attributes generated from the customers' energy consumption profiles or use the profiles directly. Numerical results demonstrate that the anomaly detection methods can correctly identify the majority of anomalous days. The anomaly detection algorithms are validated using a detailed data set that has both premise level and device level consumption data. The anomalous days can be detected and eliminated when the customers' energy consumption profiles are carefully studied and the detection models are well tuned.
[1]
Alexey Tsymbal,et al.
The problem of concept drift: definitions and related work
,
2004
.
[2]
Dorothy E. Denning,et al.
An Intrusion-Detection Model
,
1987,
IEEE Transactions on Software Engineering.
[3]
M. Zweig,et al.
Receiver-operating characteristic (ROC) plots: a fundamental evaluation tool in clinical medicine.
,
1993,
Clinical chemistry.
[4]
Eric R. Ziegel,et al.
Probability and Statistics for Engineering and the Sciences
,
2004,
Technometrics.
[5]
Jeffrey C. Lagarias,et al.
Convergence Properties of the Nelder-Mead Simplex Method in Low Dimensions
,
1998,
SIAM J. Optim..
[6]
Brendan Kirby.
Demand Response For Power System Reliability: FAQ
,
2006
.
[7]
Bernd Freisleben,et al.
CARDWATCH: a neural network based database mining system for credit card fraud detection
,
1997,
Proceedings of the IEEE/IAFE 1997 Computational Intelligence for Financial Engineering (CIFEr).
[8]
R. Balian.
Entropy, a Protean Concept
,
2004
.
[9]
B.J. Kirby.
Load Response Fundamentally Matches Power System Reliability Requirements
,
2007,
2007 IEEE Power Engineering Society General Meeting.
[10]
VARUN CHANDOLA,et al.
Anomaly detection: A survey
,
2009,
CSUR.