datawhalechina/time-series-learning

1. This repository is a notebook-based learning and solution workspace for the 2021 “AI Earth” ENSO forecasting competition. It is not a packaged application; the architecture is organized as five progressively more advanced analysis/modeling tasks, each implemented in a Jupyter notebook.

2. The problem boundary is a spatiotemporal prediction pipeline: use the past 12 months of four climate variables to predict future Nino3.4 index values for horizons 1–24 months. The repository’s core technical challenge is converting multi-year gridded climate data into a model-ready tensor representation and then forecasting a long lead-time sequence.

3. Task1, “气象数据分析常用工具,” is the tooling/onboarding layer. It is meant to establish the meteorological data analysis stack and working conventions for handling climate datasets, likely including file inspection, array handling, and domain-specific preprocessing concepts. This forms the foundation for the later tasks.

4. Task2, “数据分析,” is the data understanding and exploration layer. Its role is to inspect the provided training data structure, temporal overlap, variable semantics, and sample construction. This task is where the repository explains how the raw climate fields are organized and how the Nino3.4 target relates to the input data.

5. The dataset boundary is important to the architecture. Training data comes from CMIP5, CMIP6, and SODA in NetCDF format with dimensions like year, month, lat, lon, and four channels: SST, T300, Ua, Va. Test data is provided as .npy tensors shaped as (12, lat, lon, 4). This repo’s modeling notebooks bridge those two formats by transforming climate grids into model inputs.

6. Task3, “模型建立之 CNN+LSTM,” is the first deep learning modeling branch. It combines convolutional layers for spatial feature extraction with LSTM layers for temporal modeling, reflecting the repository’s main design pattern: separate spatial and temporal information processing, then fuse them for sequence forecasting.

7. Task4, “模型建立之 TCNN+RNN,” is an alternative model family focused on temporal convolution plus recurrent processing. It represents a second architecture path for the same prediction problem, showing the repo explores multiple ways to capture long-range dependencies while keeping the model relatively compact for the small dataset regime.

8. Task5, “模型建立之 SA-ConvLSTM,” is the most advanced modeling branch in the repo. It adds attention-aware ConvLSTM-style spatiotemporal modeling, indicating an effort to improve feature selection across both space and time and better handle the long forecast horizon.

9. The fig/ subfolders under Tasks 3–5 are supporting assets for model explanation and visualization. They likely contain architecture diagrams, loss curves, or result plots used by the notebooks, so they belong to the presentation/analysis layer rather than the data or model code path.

10. The README is effectively the project’s architecture document: it defines the competition setting, data modality, sample construction, and evaluation metric. That metric matters architecturally because it rewards both correlation and RMSE across 24 forecast steps, so the modeling stack is designed for multi-step sequence prediction rather than single-step regression.

11. Overall data flow is: raw climate reanalysis/simulation data → exploratory analysis and preprocessing understanding → spatiotemporal tensor construction → alternative deep learning models (CNN+LSTM, TCNN+RNN, SA-ConvLSTM) → 24-month Nino3.4 predictions evaluated by a competition-specific score. The repository contains no deployment or serving layer; its boundary ends at offline experimentation and model comparison.

12. The main architectural theme is iterative competition workflow rather than software system decomposition. Each notebook is a stage in the analysis/modeling pipeline, and the repo’s value is in documenting how to move from domain data understanding to model design choices under a small-sample, long-horizon forecasting constraint.