|
Abstract:
|
An efficient anomaly detection in the training data is crucial for achieving high forecasting accuracy. When dealing with a large number of time series, it is not feasible to clean each series manually, especially when model re-training is periodically required. In this talk I will present several automatic anomaly detection algorithms for big noisy data, with their applications in electricity demand forecasting. In particular, I will discuss how to remove outliers within the electricity demand context while retaining meaningful extreme values, and how to select an optimal training period.
|
