---
license: cc-by-4.0
language:
- en
tags:
- weather
- forecast
- nwp
- solar-energy
- time-series
- zarr
- open-meteo
pretty_name: APAC NWP Forecast (analysis-ready Zarr cube)
---

# APAC NWP Forecast — analysis-ready Zarr cube

Hourly numerical weather prediction (NWP) forecasts for the Asia-Pacific region
(lat −44…46, lon 92…154 — Japan, Taiwan, Southeast Asia, Australia and surrounding
seas), archived **every run** from multiple weather models. Variables are chosen for
**solar-energy applications** (irradiance, temperature, cloud cover, wind, …).

Every run is stored in **one growing, analysis-ready Zarr cube per model** with axes
`init_time × lead_time × latitude × longitude` — a record of *past forecasts*, useful for
forecast-correction / model-blending ML (not reconstructable from reanalysis). New runs are
appended daily; the `run_init` axis is pre-allocated to the end of 2027, so the cube grows in
place. **The full model grid is kept, including ocean cells.**

Data is sourced from the [Open-Meteo](https://open-meteo.com) open-data distribution and
exported with the open-source [open-meteo](https://github.com/open-meteo/open-meteo)
toolchain.

## Dataset family

| Dataset | Format | Role |
|---|---|---|
| **`jimtseng/apac-nwp-forecast`** (this one) | Zarr cube (Mode A, sharded) | **Analysis-ready Silver — use this** |
| [`jimtseng/apac-nwp-forecast-raw`](https://huggingface.co/datasets/jimtseng/apac-nwp-forecast-raw) | per-run parquet | Bronze — immutable per-run capture (from go-forward onward) |
| [`jimtseng/apac-nwp-forecast-zip`](https://huggingface.co/datasets/jimtseng/apac-nwp-forecast-zip) | per-run `.zarr.zip` | Historical cold archive (the pre–Bronze original capture) |

## Coverage at a glance

**Spatial extent** — `dwd_icon` GHI over the whole domain (−44…46°N, 92…154°E):

![ICON full APAC GHI map](assets/icon_apac_domain.png)

Example `jma_msm` variables (~5 km, Japan domain) — many fields beyond GHI:

| Cloud cover (%) | 10 m wind speed (m/s) |
|:---:|:---:|
| ![JMA cloud cover](assets/jma_cloud_cover.png) | ![JMA wind speed](assets/jma_wind_speed.png) |

## Models

| Model | Provider | Resolution | Runs/day | Horizon | Grid (incl. ocean) |
|---|---|---|---|---|---|
| `jma_msm` | JMA (Japan) | 0.0625°×0.05° (~5 km) | 8 (3-hourly) | 78 h (00/12 UTC) / 39 h (other runs) | 473×481, 22.4–47.6°N · 120–150°E |
| `dwd_icon` | DWD (Germany) | 0.125° (~11 km) | 4 | 180 h (7.5 days) | 721×497, −44…46°N · 92…154°E |

Each model is a separate cube: `jma_msm_silver.zarr`, `dwd_icon_silver.zarr`.

Notes:
- Hourly values beyond a model's native hourly range (ICON > 78 h) are interpolated by
  Open-Meteo from 3-hourly steps (solar-geometry-aware clearness-index for radiation,
  hermite otherwise). Daily totals stay reliable; sub-3-hourly cloud variability there is smoothed.
- Horizon is **run-dependent** (`jma_msm`: 00/12 UTC reach 78 h, other runs 39 h). Slots
  beyond a run's real horizon read back as the missing-value sentinel (NaN / masked).
- Ocean cells carry real model values. `elevation` is 0 over sea — use it (or `location_id`)
  to mask land/sea.

## Structure

Each `<model>_silver.zarr` is a Zarr v3 store (Mode A):

- **dims** `(run_init, lead, latitude, longitude)` — `lead` = forecast lead time in **hours**
  after `run_init` (`valid_time = run_init + lead`). `run_init` grows as new runs are appended.
- **coords** `run_init`, `lead`, `latitude`, `longitude`, plus 2-D static `elevation` and
  `location_id` `(latitude, longitude)` (stored once).
- **data variables** — each weather variable is its own array in its **native dtype**
  (uint8 / uint16 / float32).
- **sharding** — one shard per run (`run_init=1 × all-lead × whole-grid`) containing small
  inner chunks, so point/series reads touch few bytes while file counts stay HF-friendly.
- **`slot_filled`** — `(run_init,)` int8 marker: 1 = this run has data, 0 = empty future slot.

Unwritten (empty / beyond-horizon) cells read back as the sentinel (NaN for float, masked for int).

## Schema (variables)

| Variable | Type | Description |
|---|---|---|
| `shortwave_radiation_wattPerSquareMetre` | uint16 | GHI, backwards-averaged over the previous hour |
| `direct_radiation_wattPerSquareMetre` | uint16 | Direct horizontal irradiance |
| `diffuse_radiation_wattPerSquareMetre` | uint16 | Diffuse irradiance |
| `direct_normal_irradiance_wattPerSquareMetre` | uint16 | DNI (unstable at sun elevation < 5°; filter before use) |
| `temperature_2m_celsius` | float32 | 2 m air temperature |
| `relative_humidity_2m_percentage` | uint8 | 2 m relative humidity |
| `wind_speed_10m_metrePerSecond` | float32 | 10 m wind speed |
| `surface_pressure_hectopascal` | float32 | Surface pressure |
| `precipitation_millimetre` | float32 | Hourly precipitation |
| `cloud_cover_percentage` (+ `_low` / `_mid` / `_high`) | uint8 | Cloud cover layers |
| `snow_depth_metre` | float32 | Snow depth (**`dwd_icon` only**) |
| `snowfall_water_equivalent_millimetre` | float32 | Hourly snowfall, water equivalent (**`dwd_icon` only**) |
| `elevation` | float32 | Grid-cell elevation (m); 0 over sea |
| `location_id` | int32 | Grid point index within the model grid |

**Per-model variable set:** `dwd_icon` has **15** weather variables; `jma_msm` has **13**
(no snow — its upstream source provides none). The earliest `dwd_icon` runs (2026-03-19/20/21)
predate the snow addition and also lack snow (those slots are the sentinel).

Integer columns reflect the source quantization (radiation step 1 W/m², cloud/humidity integer %);
values are identical to the float representation, raw grid-cell values (no elevation downscaling).

## Usage

```python
import xarray as xr

# open the whole growing cube straight from HF (lazy; reads only the chunks you touch)
ds = xr.open_zarr(
    "hf://datasets/jimtseng/apac-nwp-forecast/dwd_icon_silver.zarr",
    consolidated=True,
)

# only the runs that actually have data (skip empty future slots)
ds = ds.isel(run_init=(ds["slot_filled"] == 1))

# a point time series: all runs × all leads at one grid point
ts = ds["shortwave_radiation_wattPerSquareMetre"].sel(
    latitude=25.0, longitude=121.5, method="nearest"
)

# "all forecasts valid at time T" is the diagonal run_init + lead == T
```

Needs a recent `zarr>=3` and `huggingface_hub`. `hf://` streaming reads shards on demand; for
heavy use, download the cube (or specific variables) locally first.

## License & attribution

Data: [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/) —
<a href="https://open-meteo.com/">Weather data by Open-Meteo.com</a>, based on open data by
JMA (MSM) and DWD (ICON). Please retain this attribution when redistributing or displaying the data.