Identifying novelties and anomalies for incremental learning in streaming time series forecasting