era5-t2m-lstm — LSTM 24 h forecast of ERA5 2 m temperature

Predicts the next 24 hours of 2 m air temperature at any ERA5 grid cell in the region below, from the previous 72 hours at that cell. One shared model for every cell; location enters through normalised latitude/longitude inputs.

Data

• Source: ERA5 reanalysis, 2m_temperature, hourly, 0.25° grid, via the Copernicus Climate Data Store. Contains modified Copernicus Climate Change Service information 2021–2025.
• Region: 41.75°N–57.0°N, 95.0°W–74.25°W (62 × 84 cells).
• Splits (chronological): train 2021-01-01 → 2023-12-31, val 2024-01-01 → 2024-12-31, test 2025-01-01 → 2025-12-31.
• Normalisation: z-score with train-only mean 276.424 K, std 13.571 K (scaler.json).

Architecture

Input (72, 7) → 2 × LSTM(128) (dropout 0.1) → Dense(24), linear. 204,312 parameters. Keras 3.13.2 / TensorFlow 2.20.0.

Test metrics (2025, every cell, t0 every 6 h, 7,525,560 windows)

lead (h)	RMSE (K)	MAE (K)	persistence RMSE (K)
1	1.505	1.194	0.879
3	1.946	1.521	2.649
6	2.578	1.962	4.453
12	3.561	2.703	6.311
18	4.072	3.082	5.735
24	4.481	3.405	4.979
overall	3.462	2.542	5.106

Skill vs persistence: 0.322 (1 − RMSE / persistence RMSE). Validation (2024): RMSE 3.256 K, skill vs persistence 0.342.

Input features (per timestep)

#	feature
0	t2m_norm
1	sin_hour
2	cos_hour
3	sin_doy
4	cos_doy
5	lat_norm
6	lon_norm

t2m_norm = (t2m_K − mean) / std; hour and day-of-year are sin/cos encoded (UTC); lat_norm = (lat − 41.75) / 15.25, lon_norm = (lon − (-95.0)) / 20.75.

Usage

import json, numpy as np, keras
from huggingface_hub import hf_hub_download

repo = "hersxy/era5-t2m-lstm"
model  = keras.saving.load_model(hf_hub_download(repo, "model.keras"), compile=False)
scaler = json.load(open(hf_hub_download(repo, "scaler.json")))
cfg    = json.load(open(hf_hub_download(repo, "config.json")))

# X: float32 [batch, 72, 7] built exactly as the feature table above
y_norm = model.predict(X)                       # [batch, 24]
y_K    = y_norm * scaler["t2m_std_K"] + scaler["t2m_mean_K"]

Limitations

• Valid only inside the training bbox and for hourly ERA5-like inputs in kelvin.
• Trained on 2021–2023; no guarantee under climate drift or extreme events outside that range.
• Univariate: uses only past temperature at the same cell plus time/location encodings — no winds, pressure, or neighbouring cells.