Welcome to PaddleTS
PaddleTS is an easy-to-use Python library for deep time series modeling, focusing on the state-of-the-art deep neural network models based on PaddlePaddle deep learning framework. It aims to provide great flexibility and excellent user experiences for practitioners and professionals. It’s featured with:
A unified data structure named TSDataset for representing time series data with one or multiple target variables and optional different kinds of covariates (e.g. known covariates, observed covariates, static covariates, etc.)
A base model class named PaddleBaseModelImpl , which inherits from the PaddleBaseModel and further encapsulates some routine procedures (e.g. data loading, callbacks setup, loss computation, training loop control, etc.) and allows developers to focus on the implementation of network architectures when developing new models
A set of state-of-the-art deep learning models containing NBEATS, NHiTS, LSTNet, TCN, Transformer, DeepAR(Probabilistic), Informer, etc. for forecasting, TS2Vec for representation
A set of transformation operators for data preprocessing (e.g. missing values/outliers handling, one-hot encoding, normalization, and automatic date/time-related covariates generation, etc.)
A set of analysis operators for quick data exploration (e.g. basic statistics and summary)
Automatic time series modeling module (AutoTS) which supports mainstream Hyper Parameter Optimization algorithms and shows significant improvement on multiple models and datasets
Third-party (e.g. scikit-learn) ML models & data transformations integration
Recently updated:
Released a new time series representation model, i.e. Contrastive Learning of Disentangled Seasonal-trend Representations(CoST)
Time series anomaly detection model supported, with three deep models released, including AE(AutoEncoder), VAE(Variational AutoEncoder), and AnomalyTransformer
Third-party pyod ML models integration supported
Support time series model ensemble with two types of ensemble forecaster, StackingEnsembleForecaster and WeightingEnsembleForecaster proposed
RNN time series forecasting model supports categorical features and static covariates
New representation forecaster to support representation models to solve time series forecasting task
Support joint training of multiple time series datasets
In the future, more advanced features will be coming, including:
More time series anomaly detection models
More time series representation learning models
More probabilistic forecasting models
Scenario-specific pipelines which aim to provide an end-to-end solution for solving real-world business problems
And more
Project GitHub: https://github.com/PaddlePaddle/PaddleTS