antontuzovAI/weatherformer-jena

WeatherFormer (Jena Climate)

Model Description

This is a Time-Series Transformer (WeatherFormer) trained to predict the next 24 hours of temperature based on the past 168 hours (1 week) of multivariate weather data.

Dataset Info

Trained on the Max Planck Jena Climate Dataset (420,000+ hourly records).

Architecture Details

• Input Features: Atmospheric Pressure, Temperature, Humidity, Wind Speed, Max Wind Speed
• Target Feature: Temperature (Celsius)
• Lookback Window: 168 hours (1 week)
• Forecast Horizon: 24 hours (1 day)
• Base Architecture: 1D Convolutional Embedding + Transformer Encoder

How to Load and Use this Model

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

repo_id = 'antontuzovAI/weatherformer-jena'
weights_path = hf_hub_download(repo_id=repo_id, filename='model_weights.pth')
mean = np.load(hf_hub_download(repo_id=repo_id, filename='mean.npy'))
std = np.load(hf_hub_download(repo_id=repo_id, filename='std.npy'))