docs / data / weather

Weather Forecast Data

Daily weather prediction files from the Meteo2 sFTP server. Each file contains a multi-group forecast for 45 Swiss weather stations, covering ~3 days ahead at 3-hour resolution.

file overview

how to read a weather file — step by step

Each file is named after the day the forecast was issued (e.g. Pred_2023-10-03.csv = forecast issued on October 3rd). Inside, every row is one predicted value for one measurement, at one weather station, at one target time, from one prediction group.

Time,Value,Prediction,Site,Measurement,Unit

// Pred 00, temperature — has real values
2023-10-03 09:00:00+00:00,18.1,00,Sion,PRED_T_2M_ctrl,°C

// Pred 00, radiation — SENTINEL (no data!)
2023-10-03 09:00:00+00:00,-99999.0,00,Sion,PRED_GLOB_ctrl,Watt/m2

// Pred 03, same hour — radiation has real value
2023-10-03 09:00:00+00:00,351.9,03,Sion,PRED_GLOB_ctrl,Watt/m2

// Pred 01, different hours (offset +1)
2023-10-03 01:00:00+00:00,13.5,01,Sion,PRED_T_2M_ctrl,°C
2023-10-03 04:00:00+00:00,13.0,01,Sion,PRED_T_2M_ctrl,°C
2023-10-03 07:00:00+00:00,14.1,01,Sion,PRED_T_2M_ctrl,°C
2023-10-03 10:00:00+00:00,21.0,01,Sion,PRED_T_2M_ctrl,°C

example 1 — one prediction group covers specific hours

Prediction group 01 for Sion, temperature — it covers hours 01, 04, 07, 10, 13, 16, 19, 22 (every 3 hours, offset by 1):

Notice: hours go 01, 04, 07, 10... — not every hour. To get hours 00, 03, 06... you need prediction group 00. To get hours 02, 05, 08... you need prediction group 02. All three together = full 24h coverage.

example 2 — multiple predictions for the same hour

For Sion at 09:00, temperature — prediction groups 00, 03, 06, 09, 12, 15 all cover this hour (they all use the 00/03/06/09... offset). Each gives a slightly different forecast:

These are different model initializations giving slightly different forecasts. The average (19.25°C) is the best estimate. The spread (18.1 to 19.6) indicates uncertainty.

example 3 — the sentinel problem in prediction 00

For Sion at 09:00, solar radiation — prediction 00 has NO DATA (-99999.0), while predictions 03 and 06 have real values:

putting it all together

CSV column structure

Each CSV has a header row followed by ~153,000 data rows. Columns are comma-separated.

measurements (4 types)

prediction numbers — multiple model runs per day

Each prediction number represents a separate model run. Runs 00-33 produce 3-hourly forecasts, while runs 34-45 produce daily values only. Different runs cover different time offsets: run 00 outputs hours 00, 03, 06..., run 01 outputs 01, 04, 07..., run 02 outputs 02, 05, 08... Together, a triplet (e.g. 00+01+02) covers every hour. Runs that share the same offset (00, 03, 06, 09...) give slightly different values for the same hours — these can be averaged for a more robust estimate.

Runs sharing the same hour offset (e.g. 00, 03, 06, ..., 33 all cover hour 00:00) are independent forecasts for the same target time. With ~12 runs per hour, averaging gives a more reliable estimate and the spread indicates uncertainty.

temporal context — prediction_date

The filename (Pred_YYYY-MM-DD.csv) is the prediction issue date — the day the forecast was generated. Each file forecasts ~3 days ahead. This means multiple files contain predictions for the same target timestamps, but issued on different days.

weather stations (45 sites)

For the apartment domotics use case, the relevant station is the one closest to Valais (e.g. Sion, Visp, Zermatt, Montana, Evionnaz). The exact station to use depends on the apartment location.

data volume

recommended processing strategy

alignment with sensor data