diff --git a/pyproject.toml b/pyproject.toml
index 336cb0f..a997dcc 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -10,6 +10,8 @@ dependencies = [
     "netcdf4>=1.6",
     "cdsapi>=0.7",
     "torch>=2.1",
+    "torchvision>=0.19",
+    "torchaudio>=2.5",
     "pytorch-lightning>=2.1",
     "xgboost>=2.0",
     "scikit-learn>=1.3",
@@ -19,4 +21,16 @@ dependencies = [
     "jupyter>=1.0",
     "tqdm>=4.66",
     "scipy>=1.11",
+    "h5netcdf>=1.8.1",
+    "h5py>=3.16.0",
 ]
+
+[tool.uv.sources]
+torch = { index = "pytorch-cu126" }
+torchvision = { index = "pytorch-cu126" }
+torchaudio = { index = "pytorch-cu126" }
+
+[[tool.uv.index]]
+name = "pytorch-cu126"
+url = "https://download.pytorch.org/whl/cu126"
+explicit = true
diff --git a/src/data/extract_zips.py b/src/data/extract_zips.py
new file mode 100644
index 0000000..03698e2
--- /dev/null
+++ b/src/data/extract_zips.py
@@ -0,0 +1,50 @@
+"""批量解压 CDS 下载的 ZIP 格式 NetCDF 文件"""
+import logging
+import zipfile
+from pathlib import Path
+from concurrent.futures import ThreadPoolExecutor
+
+from src.utils.config import DATA_RAW, CITIES
+
+logger = logging.getLogger(__name__)
+
+
+def extract_one(nc_path: Path) -> bool:
+    """解压单个 ZIP 伪装的 .nc 文件，原地替换为真实 NetCDF"""
+    if nc_path.stat().st_size == 0:
+        return False
+
+    with open(nc_path, "rb") as f:
+        header = f.read(4)
+
+    if header[:2] == b"CDF":  # 已经是真实 NetCDF
+        return True
+    if header[:2] != b"PK":  # 不是 ZIP，跳过
+        return False
+
+    try:
+        with zipfile.ZipFile(nc_path) as zf:
+            nc_names = [n for n in zf.namelist() if n.endswith(".nc")]
+            if not nc_names:
+                return False
+            data = zf.read(nc_names[0])
+        nc_path.write_bytes(data)
+        return True
+    except Exception:
+        return False
+
+
+def extract_all():
+    for city in CITIES:
+        city_dir = Path(DATA_RAW) / "era5" / city
+        files = sorted(city_dir.glob("*.nc"))
+        done = 0
+        for f in files:
+            if extract_one(f):
+                done += 1
+        logger.info("%s: %d/%d 已解压", city, done, len(files))
+
+
+if __name__ == "__main__":
+    logging.basicConfig(level=logging.INFO, format="%(asctime)s [%(levelname)s] %(message)s")
+    extract_all()
diff --git a/src/data/preprocess.py b/src/data/preprocess.py
index 683bc65..01b8e76 100644
--- a/src/data/preprocess.py
+++ b/src/data/preprocess.py
@@ -534,7 +534,7 @@ def preprocess_all() -> None:
         if not saved_feature_cols:
             saved_feature_cols = feature_cols
 
-        npz_path = DATA_PROCESSED / f"sequences_{city_key}.npz"
+        npz_path = DATA_PROCESSED / f"{city_key}_sequences.npz"
         np.savez_compressed(
             npz_path,
             X=X,
@@ -561,7 +561,7 @@ def preprocess_all() -> None:
     # 合并 NPZ
     all_X, all_y = [], []
     for city_key in CITIES:
-        npz_path = DATA_PROCESSED / f"sequences_{city_key}.npz"
+        npz_path = DATA_PROCESSED / f"{city_key}_sequences.npz"
         if npz_path.exists():
             data = np.load(npz_path, allow_pickle=True)
             all_X.append(data["X"])