零样本IOT时间序列预测与异常检测

一、概要介绍

部署方式：docker部署sidus ；

通用时序预测基础大模型：

• TimesFM3谷歌，单变量为主，多变量弱一些，更适合工业过程正常趋势预测与偏离检测（可间接通过预测残差法实现零样本判异）【不可商用】（timesFM2.5可商用但不支持多变量）
• Moirai Salesforce 工业场景较火，多变量、单变量 都支持，预训练覆盖工业时序，零样本泛华强，可预测残差计算异常分值【不可商用】
• Chronos-2 亚马逊时序基础模型，能处理多变量、协变量，预测分布及不确定性判断异常；【可商用】
• MOMENT 适用于跨设备跨测点的通用时序表征；
• TiRex xLSTM架构，轻量化时序预测【可商用】

零样本时序异常预训练ux模型：

• TimeRCD 相对上下文偏差RCD，判断当前时序与历史偏移，不是基于预测残差，工业与能源多行业基准表现强；【可商用】
• ChronosAD 基于Chronos时序大模型提取Embedding，经时序模块计算异常得分。【商用不确定】

二、部署方式

docker部署

三、对外接口

接口：/timeseries/v1/forecasts
入参：
{
  "model": "chronos-2",
  "frequency": "D",
  "horizon": 7,
  "targets": ["actual_output", "energy_consumption"],
  "quantiles": [0.1, 0.5, 0.9],
  "data": [
    {
      "timestamp": "2026-11-01",
      "actual_output": 1250,
      "energy_consumption": 83000,
      "planned_output": 1300,
      "operating_hours": 22
    },
    {
      "timestamp": "2026-11-02",
      "actual_output": 1280,
      "energy_consumption": 84500,
      "planned_output": 1320,
      "operating_hours": 23
    }
  ],
  "future_covariates": [
    {
      "timestamp": "2026-11-08",
      "planned_output": 1350
    }
  ]
}



响应出参：
{
  "request_id": "fc_001",
  "model": "chronos-2",
  "frequency": "D",
  "horizon": 7,
  "results": [
    {
      "target": "actual_output",
      "forecast": [
        {
          "timestamp": "2026-11-08",
          "value": 1305,
          "quantiles": {
            "0.1": 1240,
            "0.5": 1305,
            "0.9": 1370
          }
        }
      ]
    },
    {
      "target": "energy_consumption",
      "forecast": [
        {
          "timestamp": "2026-11-08",
          "value": 85000,
          "quantiles": {
            "0.1": 82000,
            "0.5": 85000,
            "0.9": 89000
          }
        }
      ]
    }
  ],
  "metrics": {
    "inference_ms": 125.6
  }
}

异常检测接口：

/timeseries/v1/detect

入参：
{
  "model": "time-rcd",
  "frequency": "D",
  "targets": [
    "actual_output",
    "energy_consumption"
  ],
  "data": [
    {
      "timestamp": "2026-11-01",
      "actual_output": 1250,
      "energy_consumption": 83000
    },
    {
      "timestamp": "2026-11-02",
      "actual_output": 1280,
      "energy_consumption": 84500
    },
    {
      "timestamp": "2026-11-03",
      "actual_output": 720,
      "energy_consumption": 85000
    },
    {
      "timestamp": "2026-11-04",
      "actual_output": 1310,
      "energy_consumption": 83200
    }
  ],
  "detection": {
    "threshold_method": "quantile",
    "threshold_quantile": 0.99,
    "min_consecutive_points": 1,
    "merge_gap_points": 0,
    "return_scores": true,
    "return_events": true
  }
}


出参：

{
  "request_id": "ad_20261108_001",
  "model": {
    "id": "time-rcd",
    "variant": "multi"
  },
  "frequency": "D",
  "results": [
    {
      "series_id": "default",
      "targets": [
        "actual_output",
        "energy_consumption"
      ],
      "summary": {
        "total_points": 120,
        "anomaly_points": 3,
        "anomaly_ratio": 0.025,
        "event_count": 1
      },
      "scores": [
        {
          "timestamp": "2026-11-03",
          "score": 0.94,
          "is_anomaly": true
        }
      ],
      "events": [
        {
          "event_id": "evt_001",
          "start_time": "2026-11-03",
          "end_time": "2026-11-03",
          "max_score": 0.94,
          "severity": "warning"
        }
      ]
    }
  ],
  "metrics": {
    "inference_ms": 85.2
  }
}


多产品预测：
{
  "model": "time-rcd",
  "frequency": "D",
  "targets": ["actual_output"],
  "data": [
    {
      "series_id": "factory_A",
      "timestamp": "2026-11-01",
      "actual_output": 1250
    },
    {
      "series_id": "factory_A",
      "timestamp": "2026-11-02",
      "actual_output": 1280
    },
    {
      "series_id": "factory_B",
      "timestamp": "2026-11-01",
      "actual_output": 980
    },
    {
      "series_id": "factory_B",
      "timestamp": "2026-11-02",
      "actual_output": 420
    }
  ]
}