模型战绩
9 源融合 + LightGBM Stacking 的实际表现
准确率 Precision
53.0%
预测有雨中确实下了雨
召回率 Recall
71.7%
实际降雨被捕获的比例
F1 Score
60.9%
准确率×召回率调和平均
城市胜率
4/14
P≥60% 且 R≥60% 的城市数
📈 每日趋势
| 日期 | 准确率 | 召回率 | F1 | 混淆矩阵 |
|---|---|---|---|---|
| 09-23 | 59% | 62% | 61% | TP35 FP24 FN21 |
| 09-22 | 61% | 88% | 72% | TP60 FP38 FN8 |
| 09-21 | 51% | 84% | 63% | TP76 FP73 FN15 |
| 09-20 | 44% | 67% | 53% | TP36 FP46 FN18 |
| 09-19 | 65% | 48% | 55% | TP11 FP6 FN12 |
| 09-18 | 20% | 18% | 19% | TP2 FP8 FN9 |
| 09-17 | 50% | 38% | 43% | TP3 FP3 FN5 |
🏆 数据源排名(Brier Score 越低越好)
| # | 数据源 | Brier Score | 命中率 | 虚警率 |
|---|---|---|---|---|
| #1 | ECMWF 集合(51成员) | 0.0691 | 46% | 40% |
| #2 | DWD ICON | 0.0749 | 21% | 42% |
| #3 | ECMWF AIFS 集合(AI) | 0.0759 | 43% | 46% |
| #4 | ECMWF IFS | 0.0772 | 56% | 50% |
| #5 | Open-Meteo | 0.0772 | 56% | 50% |
| #6 | WeatherAPI | 0.0881 | 34% | 65% |
| #7 | wttr.in | 0.0925 | 35% | 63% |
| #8 | JMA GSM | 0.0947 | 44% | 66% |
| #9 | 彩云天气 | 0.1017 | 26% | 62% |
| #10 | 和风天气 | 0.1039 | 50% | 62% |
| #11 | VisualCrossing | 0.1151 | 49% | 64% |
🏙 城市维度表现
三明 ✅
P=60% R=80%
40 降雨时段
上海 ✅
P=100% R=60%
15 降雨时段
东莞 ⚠️
P=48% R=61%
36 降雨时段
北京 ⚠️
P=0% R=0%
4 降雨时段
南京 ⚠️
P=25% R=33%
15 降雨时段
南平 ⚠️
P=54% R=90%
29 降雨时段
厦门 ⚠️
P=44% R=73%
15 降雨时段
宁德 ✅
P=65% R=96%
25 降雨时段
杭州 ⚠️
P=58% R=41%
17 降雨时段
泉州 ⚠️
P=30% R=50%
6 降雨时段
漳州 ⚠️
P=48% R=70%
37 降雨时段
福州 ⚠️
P=59% R=90%
21 降雨时段
莆田 ⚠️
P=37% R=78%
9 降雨时段
龙岩 ✅
P=65% R=76%
42 降雨时段
🌲 LightGBM 模型状态 已训练
最近训练: 2026-09-22
| 模型 | Precision | Recall | F1 | 训练样本 |
|---|---|---|---|---|
| 0-12h/east_coast | 49% | 84% | 62% | 4,697 |
| 0-12h/global | 47% | 95% | 63% | 21,371 |
| 0-12h/north | 39% | 95% | 55% | 6,408 |
| 0-12h/south | 62% | 90% | 73% | 4,279 |
| 0-12h/southwest | 52% | 91% | 66% | 2,134 |
| 12-36h/east_coast | 26% | 67% | 38% | 10,248 |
| 12-36h/global | 42% | 91% | 58% | 46,561 |
| 12-36h/north | 41% | 78% | 54% | 13,913 |
| 12-36h/south | 50% | 89% | 64% | 9,315 |
| 12-36h/southwest | 51% | 93% | 65% | 4,677 |
| 36-72h/global | 43% | 85% | 57% | 51,148 |
| 36-72h/north | 57% | 43% | 49% | 15,297 |
| 36-72h/south | 51% | 85% | 64% | 10,226 |
| 36-72h/southwest | 42% | 87% | 57% | 5,140 |
| precip_regression | 0% | 0% | 0% | 21,371 |
⚙️ 技术架构
9 数据源
LightGBM Stacking
Regime-Based BMA
多时距独立建模
城市分区训练
降水量回归
Platt Scaling 校准
自适应权重学习