yangwohenmai/lstm

This repository contains a set of Python tutorials focused on using a type of neural network called LSTM to predict future values in time-ordered data. LSTM stands for Long Short-Term Memory, and it is a kind of model that can learn patterns across sequences, making it well suited for data where order matters, such as sales figures over months or sensor readings over time. The README is written in Chinese and organizes the content into several groups. The first group covers single-variable forecasting using shampoo sales data, walking through steps like scaling the numbers, converting the series into a format the model can learn from, and then building and evaluating the model. A second group handles multi-variable forecasting using air pollution data, which involves predicting one value based on several related inputs recorded at the same time. A third group covers multi-step forecasting, where the model predicts several future time points at once rather than just the next one. Additional sections explore encoder-decoder LSTM patterns, stateful networks, and techniques for preparing data: handling missing time steps, normalizing values, and reshaping input arrays into the format these models expect. The project also includes worked examples using airline passenger data and a dedicated section on diagnosing whether a model is underfitting or overfitting. Each topic is broken into numbered sub-steps showing what the code does at each stage. The code is built on TensorFlow and Keras, two widely used Python libraries for building and training neural networks. No standalone application is included, the repository is a study collection of scripts and examples intended for people learning how to apply LSTM models to real forecasting problems.