fifth commits

2025-03-26 16:59:57 +08:00 · 2025-03-26 16:59:57 +08:00 · 5a26dd003a
commit 5a26dd003a
parent a613f0eed1
27 changed files with 173 additions and 201 deletions
--- a/air_quality_prediction.ipynb
+++ b/air_quality_prediction.ipynb
--- a/images/AQI-SARIMA.png
+++ b/images/AQI-SARIMA.png
--- a/images/AQI.png
+++ b/images/AQI.png
--- a/images/PCA_cumulative_explained_variance.png
+++ b/images/PCA_cumulative_explained_variance.png
--- a/images/acf_AQI.png
+++ b/images/acf_AQI.png
--- a/images/acf_CO.png
+++ b/images/acf_CO.png
--- a/images/acf_NO2.png
+++ b/images/acf_NO2.png
--- a/images/acf_O3.png
+++ b/images/acf_O3.png
--- a/images/acf_PM10.png
+++ b/images/acf_PM10.png
--- a/images/acf_PM2.5.png
+++ b/images/acf_PM2.5.png
--- a/images/acf_SO2.png
+++ b/images/acf_SO2.png
--- a/images/all_acf_subplots.png
+++ b/images/all_acf_subplots.png
--- a/images/components_heatmap.png
+++ b/images/components_heatmap.png
--- a/images/correlation_heatmap.png
+++ b/images/correlation_heatmap.png
--- a/images/hourly_AQI.png
+++ b/images/hourly_AQI.png
--- a/images/hourly_CO.png
+++ b/images/hourly_CO.png
--- a/images/hourly_NO2.png
+++ b/images/hourly_NO2.png
--- a/images/hourly_O3.png
+++ b/images/hourly_O3.png
--- a/images/hourly_PM10.png
+++ b/images/hourly_PM10.png
--- a/images/hourly_PM2.5.png
+++ b/images/hourly_PM2.5.png
--- a/images/hourly_SO2.png
+++ b/images/hourly_SO2.png
--- a/images/hourly_trends_combined.png
+++ b/images/hourly_trends_combined.png
--- a/images/xg_by_step.png
+++ b/images/xg_by_step.png
--- a/images/xg_feature_importance.png
+++ b/images/xg_feature_importance.png
--- a/readme.md
+++ b/readme.md
@ -27,7 +27,7 @@
 ### 主成分分析解读
 1. 检验指标：
   - KMO值: 0.762>0.7。
-   - 巴赫利特检验卡方值: 90424.712, p值: 0.0，显著。
+   - 巴赫利特检验卡方值: 90424.712, p值: 0.0<0.001，显著。
   - 提取5个特征值大于1的因子作为主成分，累积方差贡献率为78.89%。
   - 检验效果较好，说明数据适合主成分分析降维。
 2. 旋转载荷矩阵解读：
@ -82,13 +82,13 @@
   - 最终选择AIC最低的候选模型

 4. **评估指标**
-   - RMSE: 14.326 
-   - R-squared: 0.892
-   - MAE: 9.815
-   - 相比XGBoost模型预测精度略低，但保持时间序列特性
+   - RMSE: 11.893 
+   - R-squared: 0.932
+   - MAE: 7.744
+   - 相比XGBoost模型预测精度相近，但保持时间序列特性

 5. **预测结果可视化**
-   - 滚动预测效果图：![SARIMA预测](./images/sarima_forecast.png)
+   - 滚动预测效果图：![SARIMA预测](./images/AQI-SARIMA.png)
   - 置信区间覆盖率达到95%，实际值大部分落在预测区间内

 6. **残差分析**
@ -101,10 +101,11 @@
 1. 该模型使用历史AQI数据，并进行周期性编码和滞后特征构建（3小时粒度的滞后特征（最多7天）），作为特征工程。
 2. 每次预测均采用该时间点以前的真实数据，即每次预测均为单步预测。
 3. 使用随机搜索法参数调优。
-4. 评估指标：效果很好
+4. 评估指标：
   - RMSE: 11.815
   - R-squared: 0.929
   - MAE: 7.722
 5. 预测图：![图](./images/xg_by_step.png)
-6. 重要特征：观察图，AQI_lag_1最为重要，即该时刻的AQI主要由前1个观测时刻决定。AQI_lag_21也较为重要。day_of_week显示影响较小，但不是完全没有。
-7. 其实也做了利用递归直接预测一整个月的，效果看图就很明了了![图](./images/XG_pred_by_recursion.png)
+6. 重要特征：![图](./images/xg_feature_importance.png)AQI_lag_1最为重要，即该时刻的AQI主要由前1个观测时刻决定。AQI_lag_21也较为重要。day_of_week显示影响较小，但不是完全没有。
+7. 模型比较：与SARIMA模型相近，在单步预测的准确度上几乎不相上下。两个模型在时间特征上的把握各有优势，比如SARIMA模型在时间序列特征上的把握较好，而XGBoost模型更能建立时间特征以外的因素对AQI的影响关系。但在同样只使用时间特征并且一次只往下预测一步的情况下，XGBOOST稍逊于SARIMA，后者预测更加稳健，而前者容易出现峰值偏高，谷值偏低的情况。
+8. 其实也做了利用递归直接预测一整个月的，预测结果在2天以内尚可，而后趋于平缓，具体效果看图就很明了了。![图](./images/XG_pred_by_recursion.png)
--- a/requirements.txt
+++ b/requirements.txt
@ -0,0 +1,4 @@
+matplotlib==3.8.0
+numpy==2.2.4
+pandas==2.2.3
+seaborn==0.13.2
--- a/results/rolling_forecast_results.csv
+++ b/results/rolling_forecast_results.csv
@ -1,32 +0,0 @@
-日期,实际值,预测值,预测下限,预测上限
-2023-10-01,28.636363636363626,33.8493112926664,-47.635233636420566,115.33385622175336
-2023-10-02,49.727509469696976,38.66548623183452,-42.82536438890146,120.1563368525705
-2023-10-03,54.00378787878786,51.59866579632718,-29.89349013130868,133.09082172396305
-2023-10-04,23.787878787878775,49.52408793635019,-31.975556791081225,131.0237326637816
-2023-10-05,25.58712121212121,30.456530971610437,-51.04494428047694,111.95800622369781
-2023-10-06,35.7310606060606,35.746745423979746,-45.75528519745549,117.24877604541498
-2023-10-07,60.458333333333336,44.79351798522558,-36.708655136994636,126.29569110744579
-2023-10-08,45.109848484848484,55.19286349902602,-26.308989959303503,136.69471695735552
-2023-10-09,22.24206912878789,42.96484817061444,-38.533998688523916,124.46369502975278
-2023-10-10,29.909090909090914,30.472462505431565,-51.02484294737681,111.96976795823994
-2023-10-11,51.76337594696972,36.828726721058544,-44.65444843021167,118.31190187232876
-2023-10-12,91.55764678030303,49.28101201127273,-32.19996266583957,130.76198668838504
-2023-10-13,56.132575757575786,71.2726908163902,-10.207314064424025,152.75269569720442
-2023-10-14,25.9090909090909,49.357596859306355,-32.114598711533134,130.82979243014586
-2023-10-15,29.515151515151512,31.565845981296714,-49.897131310870776,113.02882327346421
-2023-10-16,35.871212121212125,36.71261110903076,-44.74640914144864,118.17163135951016
-2023-10-17,59.96910511363636,40.90517881218401,-40.53040748891922,122.34076511328723
-2023-10-18,31.07125946969696,54.730426193951956,-26.70243278829031,136.16328517619422
-2023-10-19,21.47028882575759,33.293474144135594,-48.1146067381603,114.70155502643149
-2023-10-20,24.37121212121211,31.755445784946758,-49.65029050692998,113.1611820768235
-2023-10-21,36.86363636363636,36.817366250812796,-44.50994068018028,118.14467318180587
-2023-10-22,51.18181818181816,42.106111359156806,-39.21739699758937,123.42961971590299
-2023-10-23,116.10771780303035,48.656996034766536,-32.543358544865384,129.85735061439846
-2023-10-24,149.18323863636374,84.37233576451568,3.17385767808355,165.57081385094781
-2023-10-25,92.82019412878783,99.46879763697234,18.271785495936854,180.66580977800783
-2023-10-26,42.073863636363654,59.3336809840356,-21.853904047755073,140.52126601582628
-2023-10-27,62.851543407869016,38.4106814311253,-42.772817201854096,119.59418006410469
-2023-10-28,97.23414971285435,60.366035283496224,-20.799260620504604,141.53133118749705
-2023-10-29,175.54887585532762,78.4999966197246,-2.6629202331887285,159.6629134726379
-2023-10-30,142.330137310606,121.36437355870153,40.20258279816625,202.5261643192368
-2023-10-31,117.67966077101673,92.66882789754769,11.508510163001674,173.82914563209368