Augmented Invertible Koopman Autoencoder for long-term time series forecasting
Authors: Anthony Frion, Lucas Drumetz, Mauro Dalla Mura, Guillaume Tochon, Abdeldjalil AISSA EL BEY
TMLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We demonstrate the relevance of the AIKAE through a series of long-term time series forecasting experiments, on satellite image time series as well as on a benchmark involving predictions based on a large lookback window of observations. We perform long-term forecasting experiments with this model in two settings. Table 1: Forecasting mean squared errors (MSEs) and mean absolute errors (MAEs) for various models and long-term forecasting tasks. |
| Researcher Affiliation | Academia | Anthony Frion EMAIL Institute of Coastal Systems Analysis and Modeling Helmholtz-Zentrum Hereon, Geesthacht, Germany, Lucas Drumetz EMAIL IMT Atlantique Lab-STICC, UMR CNRS 6285, Brest, France, Mauro Dalla Mura EMAIL Université Grenobles Alpes Grenoble INP GIPSA-lab, Grenoble, France Institut Universitaire de France, Guillaume Tochon EMAIL LRE EPITA, Le Kremlin-Bicêtre, France, Abdeldjalil Aïssa El Bey EMAIL IMT Atlantique Lab-STICC, UMR CNRS 6285, Brest, France |
| Pseudocode | No | The paper describes the methods and models using mathematical equations and prose (e.g., equations 7-12, descriptions in sections 3 and 4), but does not include any clearly labeled pseudocode or algorithm blocks. |
| Open Source Code | Yes | The codes associated to our experiments are available at https://github.com/anthony-frion/AIKAE. |
| Open Datasets | Yes | In this section, we present experiments on a set of popular long-term time series forecasting datasets. Sometimes called the "Informer benchmark" as a reference to the work of Zhou et al. (2021) that popularised it, it is comprised of the ETT datasets (including the subsets ETTh1, ETTh2, ETTm1, ETTm2), ECL, Exchange, Traffic and Weather. We work with a dataset of Sentinel-2 image time series, introduced by Frion et al. (2023b) and used as a variational data assimilation benchmark by Frion et al. (2024). |
| Dataset Splits | Yes | In Table 3, we summarize the numbers of channels, sampling periods, time series lengths (including training, validation and test subsets), and nature of the information of these datasets. From this table, one can clearly see that the ECL and Traffic datasets contain much more channels than all of the other considered datasets. Concerning the loading and separation into train, validation and test splits for each dataset, we use the code from Zeng et al. (2023), corresponding to the same settings as for the benchmarks of Wu et al. (2023) and Liu et al. (2024) that the baseline methods use. To train a KAE model as a dynamical prior, we use Ttrain = 242 time steps of data, from an area of 150 × 150 pixels. The objective is to accurately predict the Ttest = 100 steps of data that follow the window of training data, by leveraging the observations of this training window. |
| Hardware Specification | No | The paper describes the experimental setup in terms of models, datasets, and hyperparameters, but does not specify any particular hardware (e.g., GPU/CPU models, memory) used for running the experiments. |
| Software Dependencies | No | The paper mentions using the Adam algorithm for training and references model architectures like NICE and Real NVP, but does not provide specific version numbers for any software libraries, programming languages, or frameworks used in the implementation. |
| Experiment Setup | Yes | The training is performed with the Adam algorithm with a learning rate of 10-3 and momentum parameters β = (0.9, 0.999). A hyperparameter search is performed for each dataset and prediction length on the number k of coupling layers (set in the range k ∈ {3, 4}) and the width w of the hidden layer of the coupling functions (set in the range w ∈ {128, 256}). for each task we search for the best batch size among the values {4, 128, 512}. |