A Chinese tutorial progression for LSTM time-series forecasting

What it does This repository houses a structured set of educational scripts for time-series forecasting using LSTM networks via TensorFlow/Keras. It covers univariate and multivariate prediction, multi-step forecasting, and preprocessing techniques such as differencing, scaling, and handling missing values. The examples target classic datasets—shampoo sales, airline passengers, and air pollution—with a brief detour into stock-price data.

The interesting bit The material is sequenced like a course syllabus rather than a random dump of notebooks, progressing from simple one-to-one regression to stateful networks, encoder-decoder models, and stacked LSTMs. The airline-passenger section explicitly contrasts window-based, timestep-based, and memory-across-batches architectures, which is useful for understanding how framing changes model behavior.

Key highlights

• 3,658 stars indicate it is a well-used reference among Chinese-speaking learners
• Covers the full pipeline: data cleaning, feature extraction, modeling, and prediction
• Includes practical topics often skipped in beginner tutorials: overfitting/underfitting diagnosis, model save/load, and strategies for missing time steps
• Applies techniques to multiple dataset types: shampoo sales, air quality, airline passengers, and stock prices
• README and documentation are entirely in Chinese

Caveats

• This is a collection of tutorial scripts, not an installable library or framework; expect to copy and adapt rather than import
• All documentation is in Chinese, so non-Chinese readers will need translation tools
• The README contains spelling inconsistencies such as “Kerase” and “one-hotenoder”

Verdict Useful if you are learning LSTM-based forecasting and read Chinese. Skip it if you need a maintained Python package or English-first documentation.