// the find
AIStream-Peelout/flow-forecast
Deep learning PyTorch library for time series forecasting, classification, and anomaly detection (originally for flood forecasting).
Flow Forecast is a PyTorch framework for time series forecasting, classification, and anomaly detection, with roots in flood and river flow prediction. It packages 17+ architectures—from vanilla LSTM to Informer, Crossformer, and iTransformer—behind a unified JSON-config-driven training interface. Best fit for researchers or ML engineers who want a ready-made harness for benchmarking multiple architectures on the same dataset without rewriting boilerplate.
The model breadth is real: you get DA-RNN, Informer, TSMixer, Anomaly Transformer, and iTransformer all wired into the same training loop, which makes architecture comparison on your own data genuinely easy. The JSON config system means you can swap models without touching code. Test coverage is solid—there are per-model JSON configs and corresponding test files, so regressions surface quickly. W&B and GCP integrations are baked in rather than bolted on.
The package name is `flood-forecast` and the module is `flood_forecast`—the rebrand to general time series never fully landed, and that friction will bite anyone who imports it without reading the README carefully. Documentation lives across three places (ReadTheDocs, Atlassian Confluence, and a GitHub Wiki) and they're not in sync; the Confluence pages are frequently the authoritative source but the least discoverable. The config-driven approach that makes benchmarking easy makes customization painful—adding a new loss function or data loader requires finding the right registration dict rather than just subclassing. No support for probabilistic forecasting beyond a thin optional layer on GRU/LSTM; if you need proper prediction intervals, you'll be extending things yourself.