性能与用户体验全面优化

    前端性能:
    - 移除 Three.js 依赖(~600KB),天气粒子改为纯 CSS 动画 + Canvas 2D
    - Google Fonts 切换为 next/font 自托管,消除跨域字体请求
    - 合并 ScanTerminalLightTheme.module.css (37KB) 到主 CSS,亮/暗主题统一用 CSS 变量
    - 新增 /api/dashboard/init 聚合端点,首次加载 4 次往返 → 1 次
    - 添加 Service Worker 静态资源缓存,修复 PWA manifest 配置

    用户体验:
    - 新增全局错误边界 error.tsx / global-error.tsx,崩溃不再白屏
    - 决策卡和城市详情的更新时间改为相对时间("15秒前"),每秒自动刷新
    - 数据陈旧时状态标签从青色切换为琥珀色提示

    DEB 算法增强:
    - 市场扫描路径接入 Open-Meteo 多模型数据(ECMWF/GFS/ICON/JMA/HRDPS 等)
    - MAE 计算加入时间衰减(decay_factor=0.85),近期模型误差权重更高

    Scope-risk: MEDIUM — 全量 170 测试通过,前端 TypeScript/build 通过,ruff 零告警
    Tested: python -m pytest -q (170 passed), npx tsc --noEmit (0 errors), npm run build (success), ruff check .
@
This commit is contained in:
2569718930@qq.com
2026-05-14 21:31:05 +08:00
parent 37494a7192
commit 6c08a68413
21 changed files with 2062 additions and 1705 deletions
+18 -16
View File
@@ -978,10 +978,14 @@ def update_daily_record(
save_history(history_file, data)
def calculate_dynamic_weights(city_name, current_forecasts, lookback_days=7):
def calculate_dynamic_weights(city_name, current_forecasts, lookback_days=7, decay_factor=0.85):
"""
计算动态权重融合 (Dynamic Ensemble Blending, DEB)
根据过去 N 天各模型的 Mean Absolute Error (MAE) 计算倒数权重
根据过去 N 天各模型的加权 MAE(时间衰减)计算倒数权重。
- 时间衰减:越近的天误差权重越大 (decay_factor^days_ago)
- decay_factor=0.85 时,1天前权重 0.853天前 0.617天前 0.32
返回: blended_high (融合预报值), weights_info (权重展示字符串)
"""
project_root = os.path.dirname(
@@ -1001,7 +1005,6 @@ def calculate_dynamic_weights(city_name, current_forecasts, lookback_days=7):
dedup_note = "家族去重" if raw_forecast_count > len(current_forecasts) else ""
if city_name not in data or not data[city_name]:
# 没有历史数据,返回简单的平均/中位数
valid_vals = [v for v in current_forecasts.values() if v is not None]
if not valid_vals:
return None, "暂无模型数据"
@@ -1011,17 +1014,12 @@ def calculate_dynamic_weights(city_name, current_forecasts, lookback_days=7):
note = f"{note} | {dedup_note}"
return round(avg, 1), note
# 获取过去 lookback_days 天的有 actual_high 的记录
city_data = data[city_name]
sorted_dates = sorted(city_data.keys(), reverse=True)
# 我们只用真正结清(或者有比较准确最高温)的历史来算误差
# 这边简化:凡是有 actual_high 的都算进去
errors = {model: [] for model in current_forecasts.keys()}
errors: dict = {model: [] for model in current_forecasts.keys()}
days_used = 0
for date_str in sorted_dates:
# 跳过今天,今天还没出最终结果
if date_str == datetime.now().strftime("%Y-%m-%d"):
continue
@@ -1032,6 +1030,8 @@ def calculate_dynamic_weights(city_name, current_forecasts, lookback_days=7):
if actual is None:
continue
decay_weight = decay_factor ** days_used
for model in current_forecasts.keys():
if model in past_forecasts and past_forecasts[model] is not None:
try:
@@ -1039,13 +1039,12 @@ def calculate_dynamic_weights(city_name, current_forecasts, lookback_days=7):
av = float(actual)
except (TypeError, ValueError):
continue
errors[model].append(abs(pv - av))
errors[model].append((abs(pv - av), decay_weight))
days_used += 1
if days_used >= lookback_days:
break
# 如果有效历史天数 < 2 天,还是使用等权
if days_used < 2:
valid_vals = [v for v in current_forecasts.values() if v is not None]
if not valid_vals:
@@ -1056,13 +1055,16 @@ def calculate_dynamic_weights(city_name, current_forecasts, lookback_days=7):
note = f"{note} | {dedup_note}"
return round(avg, 1), note
# 计算 MAE
# 计算加权 MAE(时间衰减)
maes = {}
for model, err_list in errors.items():
if err_list:
maes[model] = sum(err_list) / len(err_list)
for model, err_weighted in errors.items():
if err_weighted:
total_weight = sum(w for _e, w in err_weighted)
if total_weight > 0:
maes[model] = sum(e * w for e, w in err_weighted) / total_weight
else:
maes[model] = sum(e for e, _w in err_weighted) / len(err_weighted)
else:
# 如果某个新模型没有历史数据,给它一个平均误差
maes[model] = 2.0
# 计算权重(用 MAE 的倒数,误差越小权重越大;加 0.1 防止除以0)