公开(公告)号：GB2633833A

公开(公告)日：2025-03-26

申请号：GB202314621

申请日：2023-09-25

Applicant: IBM

Inventor： CHRISTOPHER WATSON ,  ANDRÉS GARCIA SARAVIA ORTIZ DE MONTELLANO ,  DIETER WELLERDIEK ,  JAKOB LANG ,  GIAO NGUYEN-QUYNH

IPC: G06N20/00 ,  G06F18/211 ,  G06F18/214

Abstract: A method for time-series forecasting for time-series data with a periodic behavior larger than a respective sampling rate comprises: selecting (104) candidate time lag values from measured (102) time-series data; determining (106) first training data, e.g. feature vectors, based on time-series data and the candidate time lag values; and training (108) a first machine-learning (ML) system for building a regularized ML model, thereby determining a subset of the set of the first training data related to the most influential long-term and short-term lag values when training the regularised ML model. The method further comprises: building (110) second training data based on the long-term lag values and related measured sampled time-series data; training (112) of a second machine-learning system for time-series predictions when using measured sampled time-series data as input, the training using the set of second training data, related measured sampled time-series, the short-term lag values; wherein a first performance indicator value is indicative of a prediction performance of the first time-series machine-learning model; determining (backward optimization 114) that an element of the set of second training data is significant for the training of the first time-series machine-learning model; and determining (forward optimisation 116), that the set of second training data is complete. A table with measured sampled time-series data shifted by candidate time lag values is used for determining (106) first training data. Determining the second training data being complete comprises a Fourier transformation of an error signal between ground truth and predicted time series.