LongxingTan/Time-series-prediction

What it solves

TFTS (TensorFlow Time Series) simplifies the implementation of time series tasks, such as prediction, classification, and anomaly detection. It provides a unified interface to access both classical and state-of-the-art deep learning models, reducing the effort needed to build and experiment with complex time series architectures.

How it works

Built on TensorFlow and Keras, the library uses an AutoModel and AutoConfig system that allows users to switch between different model architectures (like Transformers or RNNs) by simply changing a model name string. It supports various data input formats, including NumPy arrays, tf.data.Dataset, and tf.keras.utils.Sequence, and provides a KerasTrainer to handle the training and saving process.

Who it’s for

It is designed for researchers, industry practitioners, and data science competitors who need to implement high-performance time series models without building every architecture from scratch.

Highlights

• Wide Model Support: Includes implementations of RNN, TCN, BERT, N-BEATS, DLinear, Seq2Seq, WaveNet, Transformer, Informer, Autoformer, and TFT.
• Flexible Data Handling: Supports both encoder-only and encoder-decoder input structures, allowing for past values and future time features.
• Task Versatility: Capable of handling forecasting, classification, and anomaly detection.
• Extensible: Allows users to build custom models by using AutoBackbone as a base and adding custom heads or embeddings.