Two-Step Meta-Learning for Time-Series Forecasting Ensemble

Vaičiukynas, Evaldas; Danėnas, Paulius; Kontrimas, Vilius; Butleris, Rimantas

doi:10.1109/access.2021.3074891

Cited by 12 publications

(4 citation statements)

References 44 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…The benchmark could motivate the research community for developing domain-independent models for the task of early forecasting. The second complication translates into an interesting research question, and could be solved by considering eTSF as a direct realization of time series compatible methods from the research areas of Meta learning [4], [11] and Transfer learning [12], [13].…”

Section: * Equal Contributionmentioning

confidence: 99%

Forecasting Early with Meta Learning

Jawed

Madhusudhanan

Yalavarthi

et al. 2023

2023 International Joint Conference on Neural Networks (IJCNN)

View full text Add to dashboard Cite

In the early observation period of a time series, there might be only a few historic observations available to learn a model. However, in cases where an existing prior set of datasets is available, Meta learning methods can be applicable. In this paper, we devise a Meta learning method that exploits samples from additional datasets and learns to augment time series through adversarial learning as an auxiliary task for the target dataset. Our model (FEML), is equipped with a shared Convolutional backbone that learns features for varying length inputs from different datasets and has dataset specific heads to forecast for different output lengths. We show that FEML can meta learn across datasets and by additionally learning on adversarial generated samples as auxiliary samples for the target dataset, it can improve the forecasting performance compared to single task learning, and various solutions adapted from Joint learning, Multi-task learning and classic forecasting baselines.

show abstract

Section: * Equal Contributionmentioning

confidence: 99%

Forecasting Early with Meta Learning

Jawed

Madhusudhanan

Yalavarthi

et al. 2023

2023 International Joint Conference on Neural Networks (IJCNN)

View full text Add to dashboard Cite

show abstract

“…As mentioned above, the anomaly detection of WTs during condition monitoring mainly depends on time series forecasting (TSF). TSF assists decision making in the foreseeing future tendency of evolution, which is generally used in scientific and engineering applications like business intelligence [12], interval forecasting [13] and price trend analysis [14]. Traditional methods like the autoregressive integrated moving average (ARIMA) model [15] and Holt-Winters seasonal method [16] pay more attention to singlevariable forecasting, while lacking the applications to complex raw time series data.…”

Section: Introductionmentioning

confidence: 99%

Anomaly Detection of Wind Turbine Driveline Based on Sequence Decomposition Interactive Network

Lyu,

He,

et al. 2023

Sensors

View full text Add to dashboard Cite

Aimed at identifying the health state of wind turbines (WTs) accurately by using the comprehensive spatio and temporal information from the supervisory control and data acquisition (SCADA) data, a novel anomaly-detection method called decomposed sequence interactive network (DSI-Net) is proposed in this paper. Firstly, a DSI-Net model is trained using preprocessed data from a healthy state. Subsequences of trend and seasonality are obtained by DSI-Net, which can dig out underlying features both in spatio and temporal dimensions through the interactive learning process. Subsequently, the trained model processes the online data and calculates the residual between true values and predicted values. To identify anomalies of the WTs, the residual and root mean square error (RMSE) are calculated and processed by exponential weighted moving average (EWMA). The proposed method is validated to be more effective than the existing models according to the control experiments.

show abstract

“…Various previous works have focused on constructing an ensemble pooling module which combines the forecasts of candidate models using either a simple or weighted average (In & Jung, 2022;Montero-Manso et al, 2020;Vaiciukynas et al, 2021;Gastinger et al, 2021;Wu & Levinson, 2021). In the cases where a weighted average is used, the weights are determined either by using a statistical model (Vaiciukynas et al, 2021) or by a trainable module (Montero-Manso et al, 2020). In either case, the weights assigned to the different members are typically fixed during the forecasting phase.…”

Section: Introduction and Related Workmentioning

confidence: 99%

Frequency-based ensemble forecasting model for time series forecasting

Saeed

2022

Comp. Appl. Math.

View full text Add to dashboard Cite

The M4 forecasting competition challenged the participants to forecast 100,000 time series with different frequencies: hourly, daily, weekly, monthly, quarterly, and yearly. These series come mainly from the economic, finance, demographics, and industrial areas. This paper describes the model used in the competition, which is a combination of statistical methods, namely auto-regressive integrated moving-average, exponential smoothing (ETS), bagged ETS, temporal hierarchical forecasting method, Box-Cox transformation, ARMA errors, Trend and Seasonal components (BATS), and Trigonometric seasonality BATS (TBATS). Forty-nine submissions were evaluated by the organizers and compared with 12 benchmarks and standards for comparison forecasting methods. Based on the results, the proposed model is listed among the 17 submissions that outperform the 12 benchmarks and standards for comparison forecasting methods, ranked 15th on average and 4th with the weekly time series. In addition, a further comparison was conducted between the proposed model and other forecasting methods on forecasting EUR/USD exchange rate and Bitcoin closing price time series. It is apparent from the results that the proposed model can produce accurate results compared to many forecasting methods.

show abstract

Two-Step Meta-Learning for Time-Series Forecasting Ensemble

Abstract: This study contains research results from European Union funded project No. J05-LVPA-K-04-0004 "Artificial intelligence and statistical methods based time series forecasting and management" where a grant was administered by Lithuanian business promotion agency.

Cited by 12 publications

References 44 publications

Forecasting Early with Meta Learning

Forecasting Early with Meta Learning

Anomaly Detection of Wind Turbine Driveline Based on Sequence Decomposition Interactive Network

Frequency-based ensemble forecasting model for time series forecasting

Contact Info

Product

Resources

About