案例库 · 案例 11 · AI 落地方案

反窃电智能诊断 —— 从「人工稽查翻表」到「模型自动锁嫌疑户」

五维特征 + 4 条硬规则 + 5 项软评分,把 65 户月度用电数据落成「高置信嫌疑 / 嫌疑 / 观察 / 正常」四级结论,每户给出命中规则与依据 · ← 返回案例库

⚠ 本页数据为仿真数据:由 build_theft_model.py 按固定随机种子(seed=20261015)生成 65 户台区样本,其中注入 15 例窃电行为(5 类手法各 3 例)用于验证模型识别能力。不是任何真实客户台账,不可作为经营成果引用。真实落地时按同样列名导入台区数据(营销系统 + 用电信息采集 + 线损台账),模型与规则不变。
65
片区总用户(户)
50
判定正常
0
观察
15
窃电嫌疑
6
其中高置信
15
注入窃电样本

一、判定模型 · 五维度特征体系

计量异常电量异常负荷异常电能质量线损

窃电不会只留一个痕迹:倒表会留反向电量,绕越分流会让电量骤降,缺相失压会同时拉高三相不平衡、压低电压合格率、抬升线损。单点异常不足以下结论,五维交叉才能定性。

维度取用字段能抓的手法
计量异常反向有功电量占比、表计开盖/失准标记反接倒表、表计篡改
电量异常月度电量环比变化率、报装容量与本期电量比值绕越分流、无表用电
负荷异常日负荷率(日均负荷 ÷ 最大负荷)表计篡改(走字偏慢)
电能质量三相电流不平衡度、电压合格率、功率因数缺相失压、绕越分流
线损台区/配变损耗率缺相失压、绕越分流

二、判定规则 · 硬规则定性,软评分定强度

🔴 硬规则(命中即定性,可单独成立)
H1 反接倒表:反向有功电量占比 > 0.5% +40
H2 用电量骤降:环比 ≤ -40% 且无检修/销户记录 +35
H3 无表用电:报装容量 > 0 但本期电量 < 历史 5% +45
H4 表计异常:开盖、失准标记 = 1 +30
🟠 软评分(加权累加,只反映异常强度,不单独定罪)
S1 日负荷率过低:<0.25 → +15;0.25~0.35 → +8
S2 三相不平衡:>20% → +15;15%~20% → +8
S3 电压合格率低:<85% → +15;85%~90% → +8
S4 线损率偏高:>8% → +10;6%~8% → +5
S5 功率因数异常:<0.80 → +8
🟢 判定逻辑
含 H1 / H3 → 窃电嫌疑·高置信 | 含 H2 / H4 或总分 ≥30 → 窃电嫌疑
总分 15~29 → 观察(列入下期重点,不动现场) | 总分 < 15 → 正常

设计取舍:硬规则负责「敢不敢报」,软评分负责「排先后」。把 S 类指标单独定罪会造成大规模误伤——老小区电压合格率天然偏低、小工业负荷率天然偏高,所以软评分只加分、不单独触发稽查。

三、检测效能(仿真样本 · 仅供参考)

15
真正例 TP(窃电被识别)
0
漏报 FN
0
误报 FP
100.0%
精确率
100.0%
召回率
100.0%
准确率
这套 100% 是构造出来的上限值,不是实战水平:样本按规则刻意可分、注入手法与规则一一对应,因此模型必然全中。真实台区的难点在「疑似但证据不足」——电量骤降可能来自停产、销户、季节性停工。上生产必须先用历史稽查结案数据回标定阈值(尤其 H2 的 -40% 与 H3 的 5%),再谈准确率。

四、各类窃电手法命中情况

窃电手法注入样本识别数命中率
无表用电33100.0%
表计篡改33100.0%
绕越分流33100.0%
反接倒表33100.0%
缺相失压33100.0%

五类手法对应五条不同证据链:反接倒表靠反向电量、绕越分流靠电量骤降、表计篡改靠开盖标记与负荷率、缺相失压靠三相不平衡 + 电压合格率 + 线损三指标同向恶化、无表用电靠报装容量与电量归零的矛盾。

五、风险评分分布(65 户,按评分降序)

高置信嫌疑 窃电嫌疑 正常 柱高 = 风险评分(满分 90+)

六、诊断明细(按风险评分降序 · 红=高置信 / 橙=嫌疑 / 黄=观察 / 绿=正常)

用户ID类型报装容量历史电量本期电量环比反向占比 日负荷率不平衡%电压合格%功率因数线损% 命中规则评分判定嫌疑手法(仿真注入)
1064居民10.8366.910.7-97.1%0.0%0.3372.298.90.964.29H2用电量骤降 | H3无表用电88窃电嫌疑·高置信是无表用电
1065居民10.2395.614.1-96.4%0.0%0.3498.798.70.8963.82H2用电量骤降 | H3无表用电88窃电嫌疑·高置信是无表用电
1057居民9.5301.1153.1-49.2%0.0%0.1494.698.00.9132.98H2用电量骤降 | H4表计异常 | S1日负荷率过低80窃电嫌疑是表计篡改
1063居民8.9409.80.1-100.0%0.0%0.5175.497.80.8945.38H2用电量骤降 | H3无表用电80窃电嫌疑·高置信是无表用电
1059工业1060.070215.351485.8-26.7%0.0%0.2387.498.00.8665.38H4表计异常 | S1日负荷率过低45窃电嫌疑是表计篡改
1056居民9.7408.3189.1-53.7%0.28%0.3498.999.80.8982.42H2用电量骤降43窃电嫌疑是绕越分流
1051工业902.651985.852060.00.1%27.89%0.6795.999.30.9433.22H1反接倒表40窃电嫌疑·高置信是反接倒表
1052商业101.02829.72508.6-11.3%26.24%0.4185.199.70.8683.94H1反接倒表40窃电嫌疑·高置信是反接倒表
1053居民10.6435.9432.4-0.8%13.2%0.4242.297.80.8714.99H1反接倒表40窃电嫌疑·高置信是反接倒表
1060居民9.8442.4309.1-30.1%0.0%0.37224.076.00.8610.03S2三相不平衡 | S3电压合格率低 | S4线损率偏高40窃电嫌疑是缺相失压
1061居民10.8366.6249.1-32.1%0.0%0.41732.281.60.8910.59S2三相不平衡 | S3电压合格率低 | S4线损率偏高40窃电嫌疑是缺相失压
1062工业1818.121152.414795.4-30.1%0.0%0.60835.774.60.8998.65S2三相不平衡 | S3电压合格率低 | S4线损率偏高40窃电嫌疑是缺相失压
1058工业894.842088.932577.3-22.6%0.37%0.2582.897.10.8663.47H4表计异常38窃电嫌疑是表计篡改
1054商业69.45276.32077.7-60.6%0.0%0.5246.799.00.8693.51H2用电量骤降35窃电嫌疑是绕越分流
1055居民11.8496.4174.5-64.8%0.0%0.3595.397.20.9453.71H2用电量骤降35窃电嫌疑是绕越分流
1047居民11.7280.8285.91.8%0.0%0.3183.198.90.9266.9—13正常否—
1009居民11.1381.0358.1-6.0%0.0%0.3283.398.20.8953.21—8正常否—
1010居民11.8225.8243.27.7%0.13%0.3237.699.80.8783.75—8正常否—
1016居民8.9230.9214.6-7.1%0.0%0.3312.497.90.9072.37—8正常否—
1018居民11.0291.5271.0-7.0%0.09%0.3098.297.90.9423.88—8正常否—
1026居民10.6252.1260.13.2%0.0%0.3242.398.80.8832.06—8正常否—
1029居民11.0423.1400.2-5.4%0.0%0.3315.098.10.9394.81—8正常否—
1036居民11.1226.9211.3-6.9%0.26%0.3399.099.30.8912.03—8正常否—
1046居民8.1268.4250.4-6.7%0.0%0.5287.598.00.9576.64—5正常否—
1048居民9.7254.6272.16.9%0.0%0.4052.098.20.9416.57—5正常否—
1049居民10.0235.9247.44.9%0.0%0.4444.697.70.8866.97—5正常否—
1050居民9.5498.3513.13.0%0.11%0.4066.299.80.8637.42—5正常否—
1001工业797.758478.461472.15.1%0.25%0.7165.098.20.9255.43—0正常否—
1002商业148.07971.08392.65.3%0.07%0.5745.599.60.8662.82—0正常否—
1003居民12.0281.4274.7-2.4%0.15%0.3987.997.10.9415.24—0正常否—
1004商业130.64551.24832.46.2%0.0%0.5426.698.50.8712.24—0正常否—
1005居民8.1332.3352.56.1%0.01%0.5365.998.70.9315.42—0正常否—
1006工业1069.694217.197471.03.5%0.04%0.6294.197.10.9113.23—0正常否—
1007居民10.4336.4348.53.6%0.0%0.484.197.70.9233.55—0正常否—
1008商业90.06549.16808.64.0%0.05%0.5916.099.70.9515.43—0正常否—
1011居民9.3462.7436.7-5.6%0.27%0.4485.997.80.8913.38—0正常否—
1012商业135.16996.76493.4-7.2%0.0%0.5398.798.10.9414.91—0正常否—
1013居民11.4471.5444.5-5.7%0.0%0.3953.297.40.9535.16—0正常否—
1014居民11.1474.7490.43.3%0.0%0.3892.498.70.8774.34—0正常否—
1015居民9.5325.6334.72.8%0.0%0.5265.798.80.9223.4—0正常否—
1017商业131.77202.56921.8-3.9%0.08%0.5083.699.20.9182.28—0正常否—
1019居民11.4262.9266.01.2%0.0%0.5436.098.10.8993.01—0正常否—
1020居民11.1372.7388.24.2%0.03%0.5177.599.80.9434.53—0正常否—
1021居民8.4493.6488.0-1.1%0.12%0.5243.698.50.8934.37—0正常否—
1022居民9.2305.9283.9-7.2%0.0%0.493.397.50.952.03—0正常否—
1023居民10.4455.1447.3-1.7%0.11%0.4993.198.80.9285.2—0正常否—
1024居民9.0431.7437.11.3%0.35%0.3753.598.90.9343.92—0正常否—
1025居民8.8402.6410.21.9%0.18%0.515.697.60.9574.93—0正常否—
1027商业189.77209.07343.81.9%0.01%0.5898.699.20.8763.72—0正常否—
1028居民9.6258.1245.8-4.8%0.25%0.3915.098.30.9594.82—0正常否—
1030商业51.92921.22836.0-2.9%0.0%0.4378.097.80.8754.26—0正常否—
1031商业107.65199.44879.6-6.2%0.0%0.5522.899.40.9023.28—0正常否—
1032居民11.6351.3361.12.8%0.01%0.4537.797.50.8922.18—0正常否—
1033商业139.95963.05576.0-6.5%0.04%0.4982.397.30.8754.83—0正常否—
1034居民11.7246.5260.85.8%0.16%0.4585.597.30.8963.91—0正常否—
1035居民9.8462.6481.24.0%0.0%0.3522.698.70.8754.74—0正常否—
1037商业164.97017.76748.9-3.8%0.0%0.5923.998.10.9414.82—0正常否—
1038居民8.4331.5348.75.2%0.06%0.5016.998.60.8934.66—0正常否—
1039居民9.2237.8247.64.1%0.0%0.5092.299.10.8813.31—0正常否—
1040商业81.16504.26555.60.8%0.0%0.6177.198.70.8822.11—0正常否—
1041工业566.276849.080990.85.4%0.29%0.8064.298.60.9582.81—0正常否—
1042商业95.57067.47159.91.3%0.01%0.594.797.50.8914.04—0正常否—
1043居民11.9482.0459.3-4.7%0.0%0.5319.098.30.9134.58—0正常否—
1044工业1868.330645.630438.9-0.7%0.01%0.672.798.90.8622.65—0正常否—
1045居民11.3345.4332.8-3.6%0.0%0.4713.397.10.9233.6—0正常否—
生成时间:2026-10-15 09:46 | 模型:反窃电智能诊断 · 硬规则 + 软评分 | 数据可复现(seed=20261015)
源文件:theft-diagnosis-data.csv ·  build_theft_model.py ·  案例说明与复用指南 ·  返回案例库