Correlated Time Series Forecasting using Modular Multi-Task Deep Neural Networks
Authors
Micu, Darius-Valer ; Cirstea, Razvan-Gabriel ; Muresan, Gabriel-Marcel
Term
4. term
Education
Publication year
2018
Submitted on
2018-06-07
Pages
16
Abstract
This thesis investigates how to improve forecasting of correlated time series using modular deep neural networks. We tackle the problem of predicting a multi-step horizon for a target series by leveraging information shared across related series. Methodologically, we introduce two variants: (1) a modular architecture in which each series is first processed by its own convolutional component (CNN) to extract local patterns, and the merged features are then fed into a recurrent neural network (RNN) for multi-step regression; and (2) an extension in which the CNN components act as auto-encoders that also reconstruct the input series, turning the system into a multi-task model that encourages more informative representations. We evaluate primarily on a real-world industrial dataset of chemical concentrations from a sewage treatment aeration tank and additionally on Google Trends data. Reported results show lower prediction errors than established linear baselines (e.g., ARIMA) and neural networks that do not exploit cross-series correlations, with benefits for both short- and long-term horizons. The models are also designed to be robust to uncorrelated input series.
Dette arbejde undersøger, hvordan korrelerede tidsserier kan forudsiges mere præcist ved hjælp af dybe, modulære neurale netværk. Vi adresserer opgaven med at forudsige et flertrins vindue for en mål-tidsserie ved at udnytte information på tværs af flere relaterede tidsserier. Metodisk introducerer vi to varianter: (1) en modulær arkitektur, hvor hver tidsserie først behandles af sin egen konvolutionelle del (CNN) for at udtrække lokale mønstre, hvorefter de sammenlagte features fødes ind i et rekurrent netværk (RNN) til flertrins-regression; og (2) en udvidelse, hvor CNN-delene fungerer som auto-encodere, der samtidig rekonstruerer inputserierne, hvilket gør modellen multi-task og tilskynder til mere informative repræsentationer. Vi anvender primært et industrielt real-world datasæt med kemiske koncentrationer fra en beluftningstank på et renseanlæg og supplerer med Google Trends-data. Evalueringen rapporterer lavere forudsigelsesfejl end både etablerede lineære modeller (fx ARIMA) og neurale netværk, der ikke udnytter korrelationer, og viser nytte for både kortsigtet og langsigtet prognose. Modellerne er desuden designet med henblik på robusthed over for ukorrelerede inputserier.
[This apstract has been generated with the help of AI directly from the project full text]
Keywords
time series ; cnn ; rnn ; forecasting ; neural networks