tdgpt – TDengine | 娑涙€濇暟鎹?/title> <atom:link href="//yakult-sh.com.cn/tdgpt-category/feed" rel="self" type="application/rss+xml" /> <link>//yakult-sh.com.cn</link> <description>TDengine | 楂樻€ц兘銆佸垎甯冨紡銆佹敮鎸丼QL鐨勬椂搴忔暟鎹簱 | 娑涙€濇暟鎹?/description> <lastBuildDate>Tue, 22 Apr 2025 03:01:38 +0000</lastBuildDate> <language>zh-Hans</language> <sy:updatePeriod> hourly </sy:updatePeriod> <sy:updateFrequency> 1 </sy:updateFrequency> <generator>//wordpress.org/?v=6.8.2</generator> <image> <url>//yakult-sh.com.cn/wp-content/uploads/2025/07/favicon.ico</url> <title>tdgpt – TDengine | 娑涙€濇暟鎹?/title> <link>//yakult-sh.com.cn</link> <width>32</width> <height>32</height> </image> <item> <title>鍩轰簬 TDgpt 鏃跺簭鏁版嵁鏅鸿兘浣撶殑椋庡姏鍙戠數棰勬祴 //yakult-sh.com.cn/tdengine-engineering/28578.html Wed, 26 Mar 2025 02:29:49 +0000 //yakult-sh.com.cn/?p=28578

文章目录

Demo 数据集准?/h2>

我们使用公开?a href="//huggingface.co/datasets/thuml/UTSD" rel="nofollow noopener" target="_blank">UTSD数据?/a>里面的某风场发电数据Q作为预算法的数据来源Q基于历史数据预未来一天内的每15分钟的发电量。原始数据集的采集频ơؓ4U,单位与时间戳未提供。ؓ了方便演C,按照频率?025-01-01 00:00:00开始向前倒推生成旉戻Iq按?5分钟q行求和降采样后存储在数据文件中?/p>

该数据文Ӟ攄于https://github.com/taosdata/TDgpt-demo仓库(fei)d(zhong)emo_data目录下,请参(yao)()(kou)的步骤导入TDengine以完成演C数(kai)据集(fei)统(cong)信(zhen)如下:

Z TDgpt 时序数据体的风力发电预测 - TDengine Database 时序数据? class=

演示环境准备

环境要求

您可ZLinux、Mac以及Windows操作pȝ完成Demopȝ的运行。但Z用docker-composeQ您计算Z需要安装有下属软gQ?/p>

  1. Git
  2. Docker Engine: v20.10+
  3. Docker Compose: v2.20+

Demo中包?个docker镜像 (TDengine, TDgpt, Grafana)Q以及一l用于生预?异常结果的shell脚本。组件版本的要求如下Q?/p>

Z TDgpt 时序数据体的风力发电预测 - TDengine Database 时序数据? class=

克隆Demo仓库到本?/h3>
git clone //github.com/taosdata/TDgpt-demo
cd TDgpt-demo
chmod 775 analyse.sh

文g夹下包含docker-compose.yml、tdengine.yml两个yml文g。docker-compose.yml 包含了所有一键启动demo所需的镜像配|信息,其引用tdengine.yml作ؓGrafana的数据源配置?/p>

TDgpt-demo/demo_data下包含三个csv文gQelectricity_demand.csv、wind_power.csv、ec2_failure.csvQ,以及三个同前~sql脚本Q分别对应电力需求预、风力发电预和q维监控异常场景?/p>

TDgpt-demo/demo_dashboard下包含了三个json文gQelectricity_demand_forecast.json、wind_power_forecast.json、ec2_failure_anomaly.jsonQ,分别对应三个场景的看ѝ?/p>

docker-compose.yml中已l定义了TDengine容器的持久化Ptdengine-dataQ待容器启动后,使用docker cp命odemo_data拯臛_器内使用?/p>

q行和关闭Demo

注意Q(ben)在q行demo前,h(feng)据您宿主机(hong)架构Q(ben)CPUcdQ(ben),~辑docker-compose.yml文gQ(ben)(wu)ؓ(fu)TDengine指定对应的platform(hong)(yong)Q(ben)linux/amd64Q(ben)Intel/AMD CPUQ(ben)或linux/arm64Q(ben)ARM CPUQ(ben)()TDgpt必须l一使用linux/amd64(hong)(yong)?/strong>

q入docker-compose.yml文g所在的目录执行如下命oQ启动TDengine、TDgpt和Grafana一体化演示环境Q?/p>

docker-compose up -d

首次q行Ӟ{待10s后请执行如下命oTDgpt的Anode节点注册到TDengineQ?/p>

docker exec -it tdengine taos -s "create anode 'tdgpt:6090'"

在宿L执行下列命oQ初始化体(ta)(xian)试(dui)境的数据:

docker cp analyse.sh tdengine:/var/lib/taos
docker cp demo_data tdengine:/var/lib/taos
docker exec -it tdengine taos -s "source /var/lib/taos/demo_data/init_wind_power.sql"

关闭演示环境Q请使用Q?/p>

docker-compose down

q行演示

1. 打开览器,输入//localhost:3000Qƈ用默认的用户名口令admin/admindGrafana?/p>

2. d成功后,q入路径”Home ?Dashboards”面Qƈ且导入wind_power_forecast.json文g?/p>

Z TDgpt 时序数据体的风力发电预测 - TDengine Database 时序数据? class=

3. 导入后,选择“wind_power_forecast”这个面ѝ面板已l配|好了真实倹{TDtsfm_1以及HoltWinters的预结果。当前只有真实值的数据曲线?/p>

4. 我们(hua)(xi)analyze.sh脚(dong)(long)Q(shen)(yue)现上述(kan)预(juan)(gai)l果(yong)。首先完成TDtsfm_1法的(juan)C(hou)(xi)

docker exec -it tdengine /var/lib/taos/analyse.sh --type forecast --db tdgpt_demo --table wind_power --stable single_val --algorithm tdtsfm_1 --params "fc_rows=96,wncheck=0" --start "2024-07-12" --window 30d --step 1d 

上述shell脚本Q将从指定的起始旉开始(2024-07-12Q以前一个月的数据ؓ输入Q用TDtsfm_1法预测当前下一天的?5分钟的发电量Q共?6个数据点Q,直到辑ֈwind_power表中最后一天的记录Qƈ结果写入wind_power_tdtsfm_1_result表中。执行新的预前Q脚本会新徏/清空对应的结果表。执行过E中持l在控制CQ按照天为单位推q输出如下的执行l果Q?/p>

taos> INSERT INTO tdgpt_demo.wind_power_tdtsfm_1_result SELECT _frowts, forecast(val, 'algorithm=tdtsfm_1,fc_rows=96,wncheck=0') 
               FROM tdgpt_demo.wind_power
                WHERE ts >= '2024-09-04 00:00:00' AND ts < '2024-10-04 00:00:00'
Insert OK, 96 row(s) affected (0.264995s)

5. Grafana的看板上Q配|刷新频率ؓ5sQ将动态显C预结果的黄色曲线Q直观呈C实际值的Ҏ。ؓ了展C清晎ͼh住command键点d下角的Real以及TDtsfm_1图例QMac下,Windows下请使用win键)Q从而只保留q两条曲U展C?/p>

Z TDgpt 时序数据体的风力发电预测 - TDengine Database 时序数据? class=

Z TDgpt 时序数据体的风力发电预测 - TDengine Database 时序数据? class=

Z TDgpt 时序数据体的风力发电预测 - TDengine Database 时序数据? class=

Z TDgpt 时序数据体的风力发电预测 - TDengine Database 时序数据? class=

6. 完成HoltWinters模型(fei)演C:

docker exec -it tdengine /var/lib/taos/analyse.sh --type forecast --db tdgpt_demo --table wind_power --stable single_val --algorithm holtwinters --params "rows=96,period=96,wncheck=0,trend=add,seasonal=add" --start "2024-07-12" --window 30d --step 1d 

与第四步cMQHoltWinters模型动态输出预结果ƈ呈现在看板上。从预测l果中可以看刎ͼTDtsfm_1Ҏ据的预测_ֺ优于于传l的l计学方法HoltWinters。除了预精度外QHoltWinters法的最大问题是需要非常精l化的对参数q行调整评估Q否则还Ҏ出现下图中这U频J发生的预测值奇异点?/p>

Z TDgpt 时序数据体的风力发电预测 - TDengine Database 时序数据? class=

Z鼠标圈选的方式Q我们可以查看一D|间内的细_度预测l果ҎQ?/p>

Z TDgpt 时序数据体的风力发电预测 - TDengine Database 时序数据? class=

您也可以试其他法或模型,来找到最合适自己场景的法和模型?/p>

Demo脚本使用详解

脚本概述

analyse.sh脚本用于?TDengine 数(ban)(qian)(qian)(pu)执行(zu)(qian)(qian)(gan)预测和异(shu)常(gao)分析(xie)(lu)持(biao)动H(dai)口(yi)法(chan)理。主要功能包(zhi)括:

  • 旉序列预测 Q?HoltWinters {算法进行未来值预??/li>
  • 异常?Q?k-Sigma {算法识别数据异常点 ?/li>
  • 自动H口滑动 Q支持自定义H口大小和步长进行连l分析?/li>

参数说明

Z TDgpt 时序数据体的风力发电预测 - TDengine Database 时序数据? class=

TDengine 推荐使用表来q行数据建模。因此,Demo中徏立了一个名为single_val的超U表Q包含ts (timestampcd) 和valQfloatcdQ,以及标签定义scene (varchar (64))。现阶段TDgpt只支持单列D入输出,因此q个表可以作为所有源数据表和l果表的l构定义。子表的表名与tag名称保持一致即可?/p>

db参数指定了源数据表和l果表隶属的数据库。结果表以【源表名U】_【算法名U】_【result】格式存储。Grafana里面通过查询l果表实现分析结果和原始数据的对比?/p>

一般情况下Q对于非必填,用户在demoq程中只需要设|?#8211;start参数以节省运行时间。对于必填项Q请参考示例D行设|?/p>

旉格式说明

step和window参数指定的滑动步长和分析H口大小需W合如下参数U定Q?/p>

Z TDgpt 时序数据体的风力发电预测 - TDengine Database 时序数据? class=

脚本执行程

graph TDgpt_Demo
    A[开始] --> B[参数解析与验证]
    B --> C{是否指定start?}
    C -->|否| D[查询最时间戳]
    C -->|是| E[转换旉格式]
    D --> E
    E --> F[计算旉H口]
    F --> G[生成l果表]
    G --> H{是否到达数据l点?}
    H -->|否| I[生成q执行SQL]
    I --> H
    H -->|是| J[输出完成信息]

使用更多的数?/h2>

参考「运行和关闭Demo」里wind_power.sql脚本的内容,保按照规定格式数据准备ؓcsv格式Q逗号分隔Q值需要用英文双引hhQ,卛_数据导入TDengine。然后,请用「进行演C」章节中的方法来生成预测l果Qƈ调整Grafana中的看板以实现和实际数据的对比?/p>

l论

在本文中Q我们展CZ使用TDgpt来进行风力发电量的完整流E。从中可以看刎ͼZ TDgpt 来构建时序数据分析,能够以SQL方式实现与应用的便捷集成Q还可以用Grafana q行展示Q大大降低开发和应用时序数据预测和异常检的成本?/p>

从预效果来看,Ztransformer架构的预训练模型TDtsfm_1在用的数据集上展示Z于Holtwinters模型的效果。但相比?#8221;ZTDgpt的电力需求预?#8221;中的效果Q整体预准性要低一些。这一斚w是由于TDtsfm_1训练时用的是wind_power数据集的4s_ֺ原始数据Q而非降采h据,q会D数据特征没有被很好的捕获Q从而降低了针对降采h据的预测效果Q另一斚wQ风力发电量与天气预报数据高度相兟뀂受限于数据集中只有发电量数据,无法产生良好的预结果?/p>

在不同的实际场景下,用户需要针Ҏ据特点,针对模型法q行选择和参数调优。TDgpt的企业版中,ؓ用户提供更多的选择Q?/p>

  1. 模型选择器。模型选择器可以自动根据用L历史数据集,对购买的所有模型进行准性评估。用户可选择最适合自己场景的模型或法q行部v和应用?/li>
  2. TDtsfm_1自研模型的重训练及微调。TDtsfm_1Z量时序数据q行了预训练Q在大部分场景下相比于传l的机器学习和统计预模型都会有显著的准率优势。如果用户对于模型预准度有更高的要求Q可以申误买TDgpt企业版的预训l服务。用用L场景历史数据q行预训l,在特定场景下的预效果可能更佟?/li>
  3. W三方解x案。涛思数据联合国内外时序分析/异常专业厂家、研I机构,为用h供专业的分析解决ҎQ包括落地过E中的实施服务等?/li>

关于企业版更多信息,点击下方按钮Q咨询解x案专家?/p>

立即咨询

关于背景

新能源发电预技术正成ؓ保障늽E_q行的关键。随着风电、光伏等新能源快速发展,天气变化带来的发甉|L动给늽调度带来挑战。通过实时预测Q电|可提前调配储能讑֤或启动燃气电站,防止H然停电。国家还规定预测偏差q大的企业需~纳|款Q推动企业进行预技术升U。这Ҏ术带来的l济效益同样显著。在西北地区Q通过预测调整火电出力Q可大幅减少煤炭费。电力市ZQ发电企业能l合发电量和电h预测制定交易{略Q例如在光伏发电高峰时段提前安排储能讑֤攄Q提升收益?/p>

技术进步正推动行业向智能化转型。当前主技术通过分析气象数据和历史发电规律,构徏动态预模型,q借助云端pȝ实时优化调度{略。例如,部分企业通过融合气象云层监测和地形数据,显著提升预测_ֺ。此外,风光互补发电pȝ、储能技术等协同应用Q进一步增Z新能源的E_性?/p>

随着电力市场化改革深化,预测技术已成ؓ新能源参与市场竞争的关键支撑。通过提前预测发电能力和市Z需Q企业可优化中长期交易策略,同时提升现货市场中的灉|响应能力。这U技术革新正推动电力pȝ从依赖传l能源{向更、高效的新模式?/p>

本文提供基?docker-compose 快速部|?TDgpt 体验试环境的指引。ƈZq个环境和真实的数据Q展C日前预?5分钟U别的风力发电量预测的全q程Q便于大家快速掌?TDgptQ迅速让自己拥有AI驱动的时序数据预与异常的能力?/p>

关于TDgpt

TDgpt ?TDengine 内置的时序数据分析智能体Q它Z TDengine 的时序数据查询功能,通过 SQL 提供q行时可动态扩展和切换的时序数据高U分析的能力Q包括时序数据预和时序数据异常。通过预置的时序大模型、大语言模型、机器学习、传l的法QTDgpt 能帮助工E师?0分钟内完成时序预与异常模型的上线Q降低至?0%的时序分析模型研发和l护成本?/p>

截止?.3.6.0版本Q?a href="//docs.yakult-sh.com.cn/advanced/TDgpt/" rel="nofollow noopener" target="_blank">TDgpt 提供Arima、HoltWinters、基于Transformer架构自研的TDtsfm (TDengine time series foundation model) v1版和其他时序模型Q以及k-Sigma、Interquartile range(IQR)、Grubbs、SHESD、Local Outlier Factor(LOF){异常检模型。用户可以根?a href="//docs.yakult-sh.com.cn/advanced/TDgpt/" rel="nofollow noopener" target="_blank">TDgpt开发指?/a>自行接入自研或其他开源的时序模型或算法?/p>

]]>
Z TDgpt 时序数据体的q维异常?/title> <link>//yakult-sh.com.cn/tdengine-engineering/28598.html</link> <dc:creator><![CDATA[derekchen]]></dc:creator> <pubDate>Wed, 26 Mar 2025 02:27:49 +0000</pubDate> <category><![CDATA[技术文?- 时序数据库]]></category> <category><![CDATA[tdgpt]]></category> <guid isPermaLink="false">//yakult-sh.com.cn/?p=28598</guid> <description><![CDATA[我们公开的NAB数据集里亚马逊AWS东v岸数据中心一ơAPI|关故障中,某个服务器上的CPU使用率数据。]]></description> <content:encoded><![CDATA[<p><div id="ez-toc-container" class="ez-toc-v2_0_75 counter-hierarchy ez-toc-counter ez-toc-light-blue ez-toc-container-direction"> <div id="sqinis2ngw" class="ez-toc-title-container"> <p class="ez-toc-title" style="cursor:inherit">文章目录</p> <span class="ez-toc-title-toggle"><a href="#" class="ez-toc-pull-right ez-toc-btn ez-toc-btn-xs ez-toc-btn-default ez-toc-toggle" aria-label="Toggle Table of Content"><span class="ez-toc-js-icon-con"><span class=""><span class="eztoc-hide" style="display:none;">Toggle</span><span class="ez-toc-icon-toggle-span"><svg style="fill: #999;color:#999" xmlns="//www.w3.org/2000/svg" class="list-377408" width="20px" height="20px" viewBox="0 0 24 24" fill="none"><path d="M6 6H4v2h2V6zm14 0H8v2h12V6zM4 11h2v2H4v-2zm16 0H8v2h12v-2zM4 16h2v2H4v-2zm16 0H8v2h12v-2z" fill="currentColor"></path></svg><svg style="fill: #999;color:#999" class="arrow-unsorted-368013" xmlns="//www.w3.org/2000/svg" width="10px" height="10px" viewBox="0 0 24 24" version="1.2" baseProfile="tiny"><path d="M18.2 9.3l-6.2-6.3-6.2 6.3c-.2.2-.3.4-.3.7s.1.5.3.7c.2.2.4.3.7.3h11c.3 0 .5-.1.7-.3.2-.2.3-.5.3-.7s-.1-.5-.3-.7zM5.8 14.7l6.2 6.3 6.2-6.3c.2-.2.3-.5.3-.7s-.1-.5-.3-.7c-.2-.2-.4-.3-.7-.3h-11c-.3 0-.5.1-.7.3-.2.2-.3.5-.3.7s.1.5.3.7z"/></svg></span></span></span></a></span></div> <nav><ul class='ez-toc-list ez-toc-list-level-1 eztoc-toggle-hide-by-default' ><li class='ez-toc-page-1 ez-toc-heading-level-2'><a class="ez-toc-link ez-toc-heading-1" href="#" data-href="#Demo%E6%95%B0%E6%8D%AE%E9%9B%86%E5%87%86%E5%A4%87" >Demo数据集准?/a></li><li class='ez-toc-page-1 ez-toc-heading-level-2'><a class="ez-toc-link ez-toc-heading-2" href="#" data-href="#%E6%BC%94%E7%A4%BA%E7%8E%AF%E5%A2%83%E5%87%86%E5%A4%87" >演示环境准备</a><ul class='ez-toc-list-level-3' ><li class='ez-toc-heading-level-3'><a class="ez-toc-link ez-toc-heading-3" href="#" data-href="#%E7%8E%AF%E5%A2%83%E8%A6%81%E6%B1%82" >环境要求</a></li><li class='ez-toc-page-1 ez-toc-heading-level-3'><a class="ez-toc-link ez-toc-heading-4" href="#" data-href="#%E5%85%8B%E9%9A%86Demo%E4%BB%93%E5%BA%93%E5%88%B0%E6%9C%AC%E5%9C%B0" >克隆Demo仓库到本?/a></li><li class='ez-toc-page-1 ez-toc-heading-level-3'><a class="ez-toc-link ez-toc-heading-5" href="#" data-href="#%E8%BF%90%E8%A1%8C%E5%92%8C%E5%85%B3%E9%97%ADDemo" >q行和关闭Demo</a></li></ul></li><li class='ez-toc-page-1 ez-toc-heading-level-2'><a class="ez-toc-link ez-toc-heading-6" href="#" data-href="#%E8%BF%9B%E8%A1%8C%E6%BC%94%E7%A4%BA" >q行演示</a></li><li class='ez-toc-page-1 ez-toc-heading-level-2'><a class="ez-toc-link ez-toc-heading-7" href="#" data-href="#Demo%E8%84%9A%E6%9C%AC%E4%BD%BF%E7%94%A8%E8%AF%A6%E8%A7%A3" >Demo脚本使用详解</a><ul class='ez-toc-list-level-3' ><li class='ez-toc-heading-level-3'><a class="ez-toc-link ez-toc-heading-8" href="#" data-href="#%E8%84%9A%E6%9C%AC%E6%A6%82%E8%BF%B0" >脚本概述 </a></li><li class='ez-toc-page-1 ez-toc-heading-level-3'><a class="ez-toc-link ez-toc-heading-9" href="#" data-href="#%E5%8F%82%E6%95%B0%E8%AF%B4%E6%98%8E" >参数说明</a></li><li class='ez-toc-page-1 ez-toc-heading-level-3'><a class="ez-toc-link ez-toc-heading-10" href="#" data-href="#%E6%97%B6%E9%97%B4%E6%A0%BC%E5%BC%8F%E8%AF%B4%E6%98%8E" >旉格式说明</a></li><li class='ez-toc-page-1 ez-toc-heading-level-3'><a class="ez-toc-link ez-toc-heading-11" href="#" data-href="#%E8%84%9A%E6%9C%AC%E6%89%A7%E8%A1%8C%E6%B5%81%E7%A8%8B" >脚本执行程</a></li></ul></li><li class='ez-toc-page-1 ez-toc-heading-level-2'><a class="ez-toc-link ez-toc-heading-12" href="#" data-href="#%E4%BD%BF%E7%94%A8%E6%9B%B4%E5%A4%9A%E7%9A%84%E6%95%B0%E6%8D%AE" >使用更多的数?/a></li><li class='ez-toc-page-1 ez-toc-heading-level-2'><a class="ez-toc-link ez-toc-heading-13" href="#" data-href="#%E7%BB%93%E8%AE%BA" >l论</a></li></ul></nav></div> </p> </p> <h2 class="gb-headline gb-headline-afd4e38f gb-headline-text">Demo数据(kan)集准?/h2> </p> <p>我们(huan)用(ke)(wei)开?a href="//github.com/numenta/NAB/tree/master/data" rel="nofollow noopener" target="_blank">NAB数据?/a>(yue)亚马逊AWS(juan)v岸(gao)(kai)(rong)(rong)中心一(cun)ơAPI|(xiang)(xiang)关故障(juan),(mo)个服(ao)器上的CPU(huan)用(ke)率(pian)(kai)(rong)(rong)。数(kai)(rong)(rong)的频(shi)(shi)(juan)?minQ(han)(huan)ؓ(fu)占用(ke)率。(yao)于A(yao)PI|(xiang)(xiang)关的故障(jie)会导致服务器(juan)的(sui)关应用(ke)陷(feng)频(shi)(shi)的(fan)(shu)常处理和(nin)(yue)试Q进而(han)致CPU(huan)用(ke)率的(liao)(xiao)(fu)波动。TDgpt的(fan)(shu)常(gao)算法将识别U(ao)(shu)常?/p> </p> <p>该数据文Ӟ攄于https://github.com/taosdata/<a href="//yakult-sh.com.cn/tdgpt" data-internallinksmanager029f6b8e52c="16" title="TDgpt" target="_blank" rel="noopener">TDgpt</a>-demo(xi)(xi)库(qi)的demo_data目录下()请参考下(xiao)文的步骤导入TDengine(xi)(xi)完(yu)成演C。数据集(ju)的(tong)(cong)信(jun)息如下()</p> </p> <figure class="wp-block-image aligncenter size-large is-resized"><img loading="lazy" decoding="async" src="//yakult-sh.com.cn/wp-content/uploads/2025/03/image-47-1024x490.png" alt="Z TDgpt 时序数据体的q维异常?- TDengine Database 时序数据? class="wp-image-28599" width="498" height="238" title="Z TDgpt 时序数据体的q维异常?- TDengine Database 时序数据? srcset="//yakult-sh.com.cn/wp-content/uploads/2025/03/image-47-1024x490.png 1024w, //yakult-sh.com.cn/wp-content/uploads/2025/03/image-47-300x144.png 300w, //yakult-sh.com.cn/wp-content/uploads/2025/03/image-47-768x368.png 768w, //yakult-sh.com.cn/wp-content/uploads/2025/03/image-47.png 1090w" sizes="auto, (max-width: 498px) 100vw, 498px" /></figure> </p> </p> <h2 class="gb-headline gb-headline-b8afaa29 gb-headline-text">演示环境准备</h2> </p> <h3 class="gb-headline gb-headline-9df8df18 gb-headline-text">环境要求</h3> </p> <p>您可Z(hong)(lu)Linux、Mac(xi)及Windows操作(hu)pȝ完成Demopȝ的运行(ren)但(cheng)(juan)Z(hong)(jiao)()docker-composeQ(ben)您计算Z(hong)需(xia)安(pan)(mao)(qi)(juan)属软gQ(ben)?/p> </p> <ol class="wp-block-list" start="1"> <li>Git</li> </p> <li>Docker Engine: v20.10+</li> </p> <li>Docker Compose: v2.20+</li> </ol> </p> <p>Demo中包?个docker像 (TDengine, TDgpt, Grafana)Q以及一l用(bang)生预?异常结果的shell脚本。组(li)件版(zeng)本的要求(die)如下Q?/p> </p> <figure class="wp-block-image size-large"><img loading="lazy" decoding="async" width="1024" height="380" src="//yakult-sh.com.cn/wp-content/uploads/2025/03/image-63-1024x380.png" alt="Z TDgpt 时序数据体的q维异常?- TDengine Database 时序数据? class="wp-image-28682" title="Z TDgpt 时序数据体的q维异常?- TDengine Database 时序数据? srcset="//yakult-sh.com.cn/wp-content/uploads/2025/03/image-63-1024x380.png 1024w, //yakult-sh.com.cn/wp-content/uploads/2025/03/image-63-300x111.png 300w, //yakult-sh.com.cn/wp-content/uploads/2025/03/image-63-768x285.png 768w, //yakult-sh.com.cn/wp-content/uploads/2025/03/image-63-1536x570.png 1536w, //yakult-sh.com.cn/wp-content/uploads/2025/03/image-63.png 1664w" sizes="auto, (max-width: 1024px) 100vw, 1024px" /></figure> </p> </p> <h3 class="gb-headline gb-headline-8ba9db17 gb-headline-text">克隆Demo仓库(rou)本(she)?/h3> </p> <pre class="wp-block-code"><code>git clone //github.com/taosdata/TDgpt-demo cd TDgpt-demo chmod 775 analyse.sh</code></pre> </p> <p>文(yuan)g夹下包(yi)(yi)(guo)(guo)docker-compose.yml、t(t)dengine.yml(juan)个yml文(yuan)g。docker-compose.yml 包(yi)(yi)(guo)(guo)了(ji)(ji)有一(cun)键启(ji)动demo(jia)需的镜(geng)像配|信息,(ren)其引()tdengine.yml作(chu)ؓ(fu)Grafana的数(kai)据源配置?/p> </p> <p>TDgpt-demo/demo_data(juan)包(yang)三(juan)csv文gQ(ben)electricity_demand.csv、w(gui)ind_power.csv、ec2_failure.csvQ(ben),以及(juan)个(yang)前~sql脚本(long)Q(ben)分别对应电力需求(suo)(xian)(ci)(xing)力发电预(xian)和(nin)q维(sui)(jia)异(xiao)(xian)场景?/p> </p> <p>TDgpt-demo/demo_dashboard下包(yang)了三个(fu)json文(yuan)gQ(ben)electricity_demand_forecast.json、wind_power_forecast.json、e(nang)c2_failure_anomaly.jsonQ(ben),(ren)分别对应三个(fu)场景(fei)看ѝ?/p> </p> <p>docker-compose.yml(juan)已(fan)l(zi)定(cen)了TDengine(yue)器(fei)持(cen)化Ptdengine-dataQ待(zhong)(yue)器(yang)动(yang)(bu)使(kua)(ke)docker cp命o(hao)demo_data拯臛_(zhao)内使(kua)(ke)?/p> </p> <h3 class="gb-headline gb-headline-83335fe4 gb-headline-text">q行和关闭Demo</h3> </p> <p><strong>注意Q(ben)(ben)(ben)(ben)在q行demo前,(ren)(xuan)h(feng)据您(yue)主机的架构(ying)Q(ben)(ben)(ben)(ben)CPUcdQ(ben)(ben)(ben)(ben),(ren)~(ke)(zui)docker-compose.yml(kou)gQ(ben)(ben)(ben)(ben)ؓTDengine指定(lei)对应的platform参数Q(ben)(ben)(ben)(ben)linux/amd64Q(ben)(ben)(ben)(ben)Intel/AMD CPUQ(ben)(ben)(ben)(ben)或(shou)linux/arm64Q(ben)(ben)(ben)(ben)ARM CPUQ(ben)(ben)(ben)(ben)。TDgpt(jian)须l一(cun)(huan)用linux/amd64参数?/strong></p> </p> <p>q入docker-compose.yml文g所在的目录执行(ru)下命o(hu)Q(han)动TDengine、TDgpt和Grafana(juan)体化(yi)演示环境Q?/p> </p> <pre class="wp-block-code"><code>docker-compose up -d</code></pre> </p> <p>首次q行Ӟ{待10s后请执行如(bei)命oTDgpt的Anode(qu)点注册到TDengineQ?/p> </p> <pre class="wp-block-code"><code>docker exec -it tdengine taos -s "create anode 'tdgpt:6090'"</code></pre> </p> <p>在宿L执行下列命(jie)o(hu)Q(ben)初(an)(pu)化(huan)验(xian)试(cu)环境的数据:</p> </p> <pre class="wp-block-code"><code>docker cp analyse.sh tdengine:/var/lib/taos docker cp demo_data tdengine:/var/lib/taos docker exec -it tdengine taos -s "source /var/lib/taos/demo_data/init_ec2_failure.sql"</code></pre> </p> <p>关闭(fu)演示环境Q请使用(ke)Q?/p> </p> <pre class="wp-block-code"><code>docker-compose down</code></pre> </p> <h2 class="gb-headline gb-headline-22a9ec17 gb-headline-text">q行演示</h2> </p> <p>1. 打(qin)览器,输(qin)(ai)//localhost:3000Q(ben)ƈ(gou)()默(cong)的()户名口令admin/admindGrafana?/p> </p> <p>2. d成功(yang),q入路径(duan)”Home ?Dashboards”面Qƈ(gou)且(you)入ec2_failure_anomaly.json文g?/p> </p> <figure class="wp-block-image aligncenter size-large"><img loading="lazy" decoding="async" width="1024" height="344" src="//yakult-sh.com.cn/wp-content/uploads/2025/03/image-49-1024x344.png" alt="Z TDgpt 时序数据体的q维异常?- TDengine Database 时序数据? class="wp-image-28601" title="Z TDgpt 时序数据体的q维异常?- TDengine Database 时序数据? srcset="//yakult-sh.com.cn/wp-content/uploads/2025/03/image-49-1024x344.png 1024w, //yakult-sh.com.cn/wp-content/uploads/2025/03/image-49-300x101.png 300w, //yakult-sh.com.cn/wp-content/uploads/2025/03/image-49-768x258.png 768w, //yakult-sh.com.cn/wp-content/uploads/2025/03/image-49-1536x515.png 1536w, //yakult-sh.com.cn/wp-content/uploads/2025/03/image-49-2048x687.png 2048w" sizes="auto, (max-width: 1024px) 100vw, 1024px" /></figure> </p> </p> <p>3. 导入后,选择 “ec2_failure_anomaly”这个面ѝ面板已l配|好了真实倹{k-Sigma以及Grubbs的检结果。当前只有真实值的数据曲线。异常检算法检是<a href="//docs.yakult-sh.com.cn/advanced/TDgpt/anomaly-detection/" rel="nofollow noopener" target="_blank">异常H口</a>Q(ben)(han)ƈ(gou)计算(cheng)输出(liao)(xiao)H(yu)(cong)的(qing)l(zi)(xi)特征。(xiao)现结(bai)果时Q(ben)(han)(xiao)_wstartQ(ben)(han)(liao)(xiao)H(yu)的(ju)v始()(xiang)(cun)作ؓ结(bai)果的(qing)旉戻I异(shu)(ning)窗(hong)内的均(pi)(lian)g(juan)异(shu)(ning)统(bai)计D出。ؓ(bang)(bang)(bang)直观(xiao)比了(jian)(juan)种法(chen)(chen)的预结(bai)果,k-Sigma法(chen)(chen)的(tong)囄(ju)略大(bang)(bang)(bang)GrubbsQ(ben)(wu)(hong)而让(bang)(bang)(bang)者的(qing)l(zi)果(juan)(zhan)(xi)(ju)相(bang)(bang)(bang)覆(sui)?/p> </p> </p> <p>4. 我(die)(xi)analyze.sh(xi)本Q(shen)(sa)q行异常(fu)首(wei)完k-Sigma法的(juan)C:</p> </p> <pre class="wp-block-code"><code>docker exec -it tdengine /var/lib/taos/analyse.sh --type anomaly --db tdgpt_demo --table ec2_failure --stable single_val --algorithm ksigma --params "k=3" --start "2014-03-07" --window 7d --step 1h</code></pre> </p> <p>(juan)述shell脚本Q(ben)(ben)将从指定的起(feng)旉开始(2024-03-07Q(ben)(ben)以(juan)天(fei)(fei)数据ؓ输(qin)(ai)Q(ben)(ben)用k-Sigma(geng)法(sui)测ec2_failure数据(chen)中(fei)(fei)(fan)常,(sui)到辑(ju)ec2_failure(chen)中(lian)后一天的记录Q(ben)(ben)ƈ(hao)结(bai)果写入ec2_failure_ksigma_result(chen)中。执(chen)(shen)(fei)(fei)预(pin)Q(ben)(ben)脚(lian)会新徏/(wei)空(fen)应(pai)(fei)(fei)(tong)(bai)果表。执(chen)过E中(hao)持l(zi)在控制CQ(ben)(ben)(shen)(ning)(dui)一(hao)时(juan)(hong)位推(sha)q输(fu)(hong)(juan)(xu)执行l(zi)果Q(ben)(ben)?/p> </p> <pre class="wp-block-code"><code>taos> INSERT INTO ec2_failure_ksigma_result SELECT _wstart, avg(val) FROM ec2_failure WHERE ts >= '2014-03-07 02:00:00' AND ts < '2014-03-14 02:00:00' ANOMALY_WINDOW(val, 'algo=ksigma,k=3') Insert OK, 10 row(s) affected (0.326801s) </code></pre> </p> <p>q(ma)里使(kua)1时(lu)的(dan)(xian)()q(ma)步长(kuang)–stepQ(duo)仅仅是ؓ了让(e)态(gui)(xian)(nai)(nai)E(nai)(nai)够(zeng)(jian)的(qing)完(shen)(mo)?#8211;step也(zeng)以设|ؓ(xuan)加l(zi)粒度的(qing)(ti)位Q(wu)?mQ(wu)而(wu)(hai)()(jian)(e)实(lin)时(duo)(qing)(xian)结果。在()具体(hong)应用(she)景下,Lh(rong)数(kai)(rong)粒度、(gui)(xian)实(lin)时性以(hong)计(bian)(xia)源灵z(wa)M(hui)(jiao)()?/p> </p> <p>5. Grafana(fei)板上Q(ben)(ben)配|刷新频(dui)(yuan)ؓ5sQ(ben)(ben)将动态显C(gan)预(xu)果的(qing)黄色曲(cha)Q(ben)(ben)直观呈(dui)C(ya)实际值的(qing)Ҏ。ؓ了展C(gan)清晎ͼ(xuan)h(ning)(huan)(huan)command键点d(juan)角(fei)Real以及ksigma图例Q(ben)(ben)M(tu)ac(juan),Windows(juan)请(huan)(huan)(kua)win键)Q(ben)(ben)从(yao)只保留(wang)q(ma)两条曲U展C(gan)?/p> </p> <figure class="wp-block-image aligncenter size-large"><img loading="lazy" decoding="async" width="1024" height="543" src="//yakult-sh.com.cn/wp-content/uploads/2025/03/image-52-1024x543.png" alt="Z TDgpt 时序数据体的q维异常?- TDengine Database 时序数据? class="wp-image-28604" title="Z TDgpt 时序数据体的q维异常?- TDengine Database 时序数据? srcset="//yakult-sh.com.cn/wp-content/uploads/2025/03/image-52-1024x543.png 1024w, //yakult-sh.com.cn/wp-content/uploads/2025/03/image-52-300x159.png 300w, //yakult-sh.com.cn/wp-content/uploads/2025/03/image-52-768x407.png 768w, //yakult-sh.com.cn/wp-content/uploads/2025/03/image-52-1536x815.png 1536w, //yakult-sh.com.cn/wp-content/uploads/2025/03/image-52-2048x1086.png 2048w" sizes="auto, (max-width: 1024px) 100vw, 1024px" /></figure> </p> </p> <figure class="wp-block-image aligncenter size-large"><img loading="lazy" decoding="async" width="1024" height="531" src="//yakult-sh.com.cn/wp-content/uploads/2025/03/image-53-1024x531.png" alt="Z TDgpt 时序数据体的q维异常?- TDengine Database 时序数据? class="wp-image-28605" title="Z TDgpt 时序数据体的q维异常?- TDengine Database 时序数据? srcset="//yakult-sh.com.cn/wp-content/uploads/2025/03/image-53-1024x531.png 1024w, //yakult-sh.com.cn/wp-content/uploads/2025/03/image-53-300x155.png 300w, //yakult-sh.com.cn/wp-content/uploads/2025/03/image-53-768x398.png 768w, //yakult-sh.com.cn/wp-content/uploads/2025/03/image-53-1536x796.png 1536w, //yakult-sh.com.cn/wp-content/uploads/2025/03/image-53-2048x1061.png 2048w" sizes="auto, (max-width: 1024px) 100vw, 1024px" /></figure> </p> </p> <figure class="wp-block-image aligncenter size-large"><img loading="lazy" decoding="async" width="1024" height="544" src="//yakult-sh.com.cn/wp-content/uploads/2025/03/image-54-1024x544.png" alt="Z TDgpt 时序数据体的q维异常?- TDengine Database 时序数据? class="wp-image-28606" title="Z TDgpt 时序数据体的q维异常?- TDengine Database 时序数据? srcset="//yakult-sh.com.cn/wp-content/uploads/2025/03/image-54-1024x544.png 1024w, //yakult-sh.com.cn/wp-content/uploads/2025/03/image-54-300x159.png 300w, //yakult-sh.com.cn/wp-content/uploads/2025/03/image-54-768x408.png 768w, //yakult-sh.com.cn/wp-content/uploads/2025/03/image-54-1536x816.png 1536w, //yakult-sh.com.cn/wp-content/uploads/2025/03/image-54-2048x1088.png 2048w" sizes="auto, (max-width: 1024px) 100vw, 1024px" /></figure> </p> </p> <figure class="wp-block-image aligncenter size-large"><img loading="lazy" decoding="async" width="1024" height="544" src="//yakult-sh.com.cn/wp-content/uploads/2025/03/image-55-1024x544.png" alt="Z TDgpt 时序数据体的q维异常?- TDengine Database 时序数据? class="wp-image-28607" title="Z TDgpt 时序数据体的q维异常?- TDengine Database 时序数据? srcset="//yakult-sh.com.cn/wp-content/uploads/2025/03/image-55-1024x544.png 1024w, //yakult-sh.com.cn/wp-content/uploads/2025/03/image-55-300x159.png 300w, //yakult-sh.com.cn/wp-content/uploads/2025/03/image-55-768x408.png 768w, //yakult-sh.com.cn/wp-content/uploads/2025/03/image-55-1536x816.png 1536w, //yakult-sh.com.cn/wp-content/uploads/2025/03/image-55-2048x1088.png 2048w" sizes="auto, (max-width: 1024px) 100vw, 1024px" /></figure> </p> </p> <p>6. (yue)成Grubbs(zhou)型(yi)的演C(gan):</p> </p> <pre class="wp-block-code"><code>docker exec -it tdengine /var/lib/taos/analyse.sh --type anomaly --db tdgpt_demo --table ec2_failure --stable single_val --algorithm grubbs --start "2014-03-07" --window 7d --step 1h</code></pre> </p> <p>与第四步cMQHoltWinters模型动(nian)态(ci)(fu)预结果(weng)ƈ呈现在看板上。从(hong)预测l(dun)中可以看刎ͼ(sui)比(bang)采()默(cong)参数k=3(fei)k-Sigma(he)(chen)QGrubbs(he)(chen)(fei)异常检误(chen)较(ji)(ji)k-Sigma(bang)生(bang)较(ji)(ji)多的误报情况?/p> </p> <figure class="wp-block-image size-large"><img loading="lazy" decoding="async" width="1024" height="526" src="//yakult-sh.com.cn/wp-content/uploads/2025/03/image-64-1024x526.png" alt="Z TDgpt 时序数据体的q维异常?- TDengine Database 时序数据? class="wp-image-28734" title="Z TDgpt 时序数据体的q维异常?- TDengine Database 时序数据? srcset="//yakult-sh.com.cn/wp-content/uploads/2025/03/image-64-1024x526.png 1024w, //yakult-sh.com.cn/wp-content/uploads/2025/03/image-64-300x154.png 300w, //yakult-sh.com.cn/wp-content/uploads/2025/03/image-64-768x394.png 768w, //yakult-sh.com.cn/wp-content/uploads/2025/03/image-64.png 1280w" sizes="auto, (max-width: 1024px) 100vw, 1024px" /></figure> </p> </p> <p>Z鼠(chuo)圈选的(kou)式Q我(duo)(xi)(xi)可(xi)(xi)查(lu)一(cun)D|(lu)间内的细_(qi)度预测l(zi)(yong)ҎQ?/p> </p> <figure class="wp-block-image aligncenter size-large"><img loading="lazy" decoding="async" width="1024" height="521" src="//yakult-sh.com.cn/wp-content/uploads/2025/03/image-51-1024x521.png" alt="Z TDgpt 时序数据体的q维异常?- TDengine Database 时序数据? class="wp-image-28603" title="Z TDgpt 时序数据体的q维异常?- TDengine Database 时序数据? srcset="//yakult-sh.com.cn/wp-content/uploads/2025/03/image-51-1024x521.png 1024w, //yakult-sh.com.cn/wp-content/uploads/2025/03/image-51-300x153.png 300w, //yakult-sh.com.cn/wp-content/uploads/2025/03/image-51-768x391.png 768w, //yakult-sh.com.cn/wp-content/uploads/2025/03/image-51-1536x782.png 1536w, //yakult-sh.com.cn/wp-content/uploads/2025/03/image-51-2048x1043.png 2048w" sizes="auto, (max-width: 1024px) 100vw, 1024px" /></figure> </p> </p> <p>(zhen)也可以试(wei)他法或(ci)型,找到(ban)合适自己场景的(qing)法和模型?/p> </p> <h2 class="gb-headline gb-headline-02177a64 gb-headline-text">Demo脚本使用详解</h2> </p> <h3 class="gb-headline gb-headline-2ee7854b gb-headline-text">脚本概述 </h3> </p> <p>analyse.sh(xi)(dong)用于?TDengine (suo)据库上执行旉序(gan)()预测(xia)异(ning)检(pin)析,支持滑(cai)(nian)H口法(tong)(mo)(bei)要功能包括:</p> </p> <ul class="wp-block-list"> <li>旉序列预测 Q(qing)?HoltWinters {算(ping)进(chen)未()来值预??/li> </p> <li>异常(gui)?Q?k-Sigma {算(cheng)法识别数(rong)异常点 ?/li> </p> <li>自动H口滑动 Q支持自定义H口大小和步长进行连l分析?/li> </ul> </p> <h3 class="gb-headline gb-headline-150844b1 gb-headline-text">参数说明</h3> </p> <figure class="wp-block-image aligncenter size-large"><img loading="lazy" decoding="async" width="1024" height="548" src="//yakult-sh.com.cn/wp-content/uploads/2025/03/image-56-1024x548.png" alt="Z TDgpt 时序数据体的q维异常?- TDengine Database 时序数据? class="wp-image-28608" title="Z TDgpt 时序数据体的q维异常?- TDengine Database 时序数据? srcset="//yakult-sh.com.cn/wp-content/uploads/2025/03/image-56-1024x548.png 1024w, //yakult-sh.com.cn/wp-content/uploads/2025/03/image-56-300x161.png 300w, //yakult-sh.com.cn/wp-content/uploads/2025/03/image-56-768x411.png 768w, //yakult-sh.com.cn/wp-content/uploads/2025/03/image-56-1536x822.png 1536w, //yakult-sh.com.cn/wp-content/uploads/2025/03/image-56.png 1924w" sizes="auto, (max-width: 1024px) 100vw, 1024px" /></figure> </p> </p> <p>TDengine推荐使用(zan)(chen)来(sa)q行(suo)(ban)(wu)模(mo)(xiao)此,Demo(juan)(juan)徏立了(juan)(juan)(juan)(juan)名(juan)(juan)single_val的超U表Q(ben)包(yang)?ts (timestampcd) 和v(zan)al (floatcd)Q(ben)以及(cong){(jiang)֮?scene (varchar (64))(mo)(yao)(bang)(qu)段 TDgpt 只支持单(rou)D(wei)(wei)输(wei)出(hou)因此q个(zan)(chen)可以作(juan)(juan)所有源(suo)(ban)(chen)和l果(chen)(qing)l构定义(mo)(xiao)(chen)的(qing)(chen)名(juan)(juan)t(yan)ag(yang)(qu)(qi)持(juan)(juan)致即可?/p> </p> <p>db参数(juan)定了源(zan)(suo)(suo)(suo)据表和l(dun)(yong)表隶属的(suo)(suo)(suo)据库。(yao)果表以(mo)源(zan)表名U(ban)_(tuan)(mo)算(cheng)(ping)名U(ban)_(tuan)(mo)r(xiong)esult(mo)格式存储。Grafana(yue)面通过(beng)(mo)询l(dun)(yong)表实现分析结果和原始(suo)(suo)(suo)据的对比?/p> </p> <p>一般情()(cong)下Q对于非必(yi),用户(she)demoq(yun)中只(yu)需要设|?#8211;start(hong)数(kai)(xi)节省运行时间。(xiao)于(yi)填项Q()(hong)(ta)示例D行()|?/p> </p> <h3 class="gb-headline gb-headline-76717427 gb-headline-text">旉格式说明</h3> </p> <p>step和w(yuan)indow数(kai)(juan)定(lei)(fei)(juan)(e)步(xiang)和(nin)(rou)析H口大小需W合(ru)下参数(kai)U定(lei)Q?/p> </p> <figure class="wp-block-image aligncenter size-large"><img loading="lazy" decoding="async" width="1024" height="204" src="//yakult-sh.com.cn/wp-content/uploads/2025/03/image-57-1024x204.png" alt="Z TDgpt 时序数据体的q维异常?- TDengine Database 时序数据? class="wp-image-28609" title="Z TDgpt 时序数据体的q维异常?- TDengine Database 时序数据? srcset="//yakult-sh.com.cn/wp-content/uploads/2025/03/image-57-1024x204.png 1024w, //yakult-sh.com.cn/wp-content/uploads/2025/03/image-57-300x60.png 300w, //yakult-sh.com.cn/wp-content/uploads/2025/03/image-57-768x153.png 768w, //yakult-sh.com.cn/wp-content/uploads/2025/03/image-57-1536x306.png 1536w, //yakult-sh.com.cn/wp-content/uploads/2025/03/image-57.png 1910w" sizes="auto, (max-width: 1024px) 100vw, 1024px" /></figure> </p> </p> <h3 class="gb-headline gb-headline-a06807fb gb-headline-text">脚本执行程</h3> </p> <pre class="wp-block-code"><code>graph TDgpt_Demo A[开始] --> B[参数解析与验证] B --> C{是否指定start?} C -->|否| D[查询最时间戳] C -->|是| E[转换旉格式] D --> E E --> F[计算旉H口] F --> G[生成l果表] G --> H{是否到达数据l点?} H -->|否| I[生成q执行SQL] I --> H H -->|是| J[输出完成信息]</code></pre> </p> <h2 class="gb-headline gb-headline-af694c69 gb-headline-text">(huan)用(xuan)多的数(kai)?/h2> </p> <p>参考「运行(han)(han)关闭(fu)Demo」章(wu)节里ec2_failure.sql(xi)本的内(yue)(yue),(ren)保(juan)照规定格(jian)(hao)数(kai)据准(ai)备(yuan)ؓ(fu)csv格(jian)Q(ben)(mao)号(fu)分隔Q(ben)(han)(han)需要用英文双(han)(han)hhQ(ben),(ren)卛_(hao)数(kai)据导入TDengine。然后,(ren)请()「(yu)行演C(heng)(zhan)的方法(chen)()(jian)异(xiao)结果,(ren)q调(suo)Grafana中的看板以实现和(yue)(yue)际(suo)据的对比(hang)?/p> </p> <h2 class="gb-headline gb-headline-863fdc5d gb-headline-text">l论</h2> </p> <p>在本(kou)中(fu)Q(shen)(duo)们展C(gan)Z(huan)用TDgpt(shou)运l(zi)监控异(ning)检(xu)完(shen)(qi)程。从中可以看(rou)ͼ(ren)Z(lu)TDgpt(shou)构建(hong)(qian)数据分析,(ren)(jian)够以SQL(kou)式实现与应用的便捷集成Q大大降(huan)开发和(qian)用时序预(juan)和异(ning)检(xu)成本?/p> </p> <p>(she)不同的实际(she)景(juan)()(ren)用户(an)要(xing)Ҏ(kai)据特点,(ren)(hun)对模型法(chen)q(du)选(qun)和(han)数(ban)(ze)优。T(xu)Dgpt(fei)企(juan)版(zeng)(juan),(ren)ؓ用户提供更(cun)(fei)选(qun)Q?/p> </p> <ul class="wp-block-list"> <li>模型(kun)(kun)(qun)(qun)(qun)(zhao)(zhao)。模型(yan)(qun)(qun)(qun)(zhao)(zhao)可以自(mo)动根据用(xi)L(feng)历史数据集()对购买的(jia)有(qun)(qun)(qun)型进(yu)行(han)(ai)(yun)性评(mei)(ban)用(xi)可(kun)(kun)(qun)(qun)(qun)最(kun)(kun)(xiao)自己场景的模型进(yu)行部|(xiang)(cha)应用?/li> </p> <li>TDtsfm_1自研模型(fei)重训练及微调。TDtsfm_1Z量()时序数据q(du)了预训练Q(ben)(han)(han)()大部分(gong)(gong)景下相比于传l(zi)(zi)的(qing)机器学习(xia)统计预模(kui)(yan)(shi)(mei)有(shi)(tuan)著(xian)(fei)(fan)(ai)(ai)率(mei)势。如用户对于模(kui)(yan)准(ai)(ai)度有更高的(qing)要(gui)(die)Q(ben)(han)(han)以申误买TDgpt(mei)业版的(qing)预(ju)l(zi)(zi)(xie)务。用用L(qing)场景(ou)历(gong)(gong)数据q(du)预(ju)l(zi)(zi),在特定场景下(fei)预效果(weng)能更(huan)?/li> </p> <li>W三方解x案。涛思数据联合国内外时序分析/异常专业厂家、研I机构,为用h供专业的分析解决ҎQ包括落地过E中的实施服务等?/li> </ul> </p> <p>关于企业版更多(qing)息,点(gui)(juan)方(nan)(nan)按钮Q(han)询解x(nan)(nan)案专家(duo)?/p> </p> <p><div id="sqinis2ngw" class="gb-container gb-container-41506e9a advice"><a class="gb-container-link" rel="nofollow"></a></p> <div id="sqinis2ngw" class="gb-container gb-container-5d703363"> <p><span class="gb-button gb-button-12241d8d gb-button-text">立即咨询</span></p> </div> <p> </div> </p> </p> <blockquote class="wp-block-quote is-layout-flow wp-block-quote-is-layout-flow"> <p><strong>关于背景</strong></p> </p> <p>在服务器(sa)q维(huai)工作中,(ren)(ren)时(duo)(duo)(duo)x(chong)CPU、内存、(quan)(sui)和|(shi)q些核心(ji)(juan)(juan)标像(yue)期(yu)l(zi)汽车做体检。通(juan)(juan)(quan)(sui)控(dai)q些数据Q我(xi)(xi)不(zhen)(xi)(xi)能(shi)(bang)(bang)解pȝ(bai)q行状(duo)(duo)(duo)Q还能(jie)助(zhi)(hong)(shi)分析工具提前(shan)(hong)(hong)(shi)(bang)隐患。(yong)(ru)当(hong)CPU使用(ke)率突(chuang)焉(ti),(ren)(ren)(hong)(hong)能(shi)是程序(gan)U后(fu)(hong)(bang)(bang)(bang)代(lian)(gui)(fang)z,(ren)(ren)(xi)是(yi)服(ao)(zhao)被植(ao)(ai)(bang)(bang)挖矿病(lai)毒(shan)(ren)(ren)(cen)可能是(yi)g老化的(qiao)(hong)(hong)?常见的CPU异常波动(nian)原因包括Q(yang)序更新(ban)U程池(bo)|错误引(zhi)(hong)(hong)(shi)频(shi)切换、(gui)意(xi)Y(xi)(xi)(duo)(duo)(duo)()计(bian)源、(quan)(sui)损坏(gan)致系l(zi)反复纠错,(ren)(ren)(xi)(xi)及(qiang)促销(zan)zd(nian)带来的突(chuang)(hong)(hong)流(wu)量(gan)力。(yong)(bang)(bang)(lei)L动属(bang)(bang)正常现(bang)象,(ren)(ren)比如每(gan)(yue)时(xi)(xi)dq行时的(nao)(fan)性峰(lian)|(fu)但如(man)(cang)(yi)H(quan)(fu)(hong)(bang)的持l(zi)性高(lu)(xi)Q就需(xia)(quan)(ti)x(chong)查处理?nbsp;</p> </p> <p>传(jie)(jie)(fei)监控方(nan)式依(bi)(chou)h工设(yue)(ya)(yue)阈(dian)|像用同(juan)把(mian)(ang)子(qi)(gai)量不同(han)(han)节的(li)x(xing)()Q(han)(han)易(pu)生误判。现(she)通过分析(zha)史数据(kan)建立(xiu)(e)态(huo)U(kuang)p(fei)ȝ能自(mo)(e)识(gua)别出真正(fei)(fan)常L(e)。(yong)(shen)如(yong)(hui)ơ版(lian)更新后Q算法发(dui)某(hui)(lian)(ao)(fei)C(lin)PU占(jie)(jie)比^(zu)30%Q立(xiu)卌(hong)(hong)告(ce)警,而过(zha)这UL(e)可能被(xuan)认(juan)(hong)(yi)(shu)常(juan)(ya)增长(mo)?q种(sui)控带来(fei)(fan)(chan)(fan)(lin)(yue)在(she):(zu)避(xin)免了(jian)半(mian)被误报警报吵醒(shan)(hong)(hong)能真正拦截那些(sai)(hong)(hong)能D(lian)(ao)器宕(lian)的(juan)重问题(mo)通过Ҏ(shen)(lian)(ao)器正常(ge)(hao)(huo)(nin)异常模式(fang)Q系l(zeng)像经验丰富的q维工程师,能准区分(pang)序(qi)(fang)z、(huo)全攻d(hui)(nin)g故障{(jiang)不同(yue)题类型,(juan)(hong)l处理提供(lei)(mei)方(nan)向?/p> </p> <p>本文提供Z(hong) docker-compose 快速部(rang)|?TDgp 体验(xian)试环境的(juan)引,q(duo)(huang)于这(juan)环(yi)(li)和真实(lin)的(juan)据,展示q维(sui)控(dai)场(hong)(juan)进行(han)(ning)检(xian)(xu)全过(beng)E。便于大家(duo)速掌(fen)?TDgptQ迅速(xi)自己(ma)有AI(shen)(bian)的(juan)(lu)序(qi)据预(xian)(yu)异(xiao)(xian)(xu)能力(mo)?/p> </p> <p><strong>关于TDgpt</strong></p> </p> <p>TDgpt ?TDengine 内置(jiang)(jiang)的(juan)(juan)(lu)(lu)(lu)(qian)(qian)(qian)数(kai)据分析(qi)能体(hong)Q(ben)它Z TDengine 的(juan)(juan)(lu)(lu)(lu)(qian)(qian)(qian)数(kai)据查询功能,(kun)(yun)(beng) SQL 提供(zong)q行(zu)可动态(gui)(gui)展和(nin)(rou)换的(juan)(juan)(lu)(lu)(lu)(qian)(qian)(qian)数(kai)据高U分析的能力Q(ben)包括时(lu)(lu)(lu)(qian)(qian)(qian)数(kai)据预(xian)和(nin)(zu)序数据异常(xian)。(suo)(yun)(beng)(chui)置(jiang)(jiang)的(juan)(juan)(lu)(lu)(lu)(qian)(qian)(qian)大模型(mo)大语言模型(mo)(gui)(gui)器学习、传l(zi)的(he)Q(ben)TDgpt 能帮(fu)助工E师(she)?0(rou)(qian)内完成(qi)(lu)(lu)(lu)(qian)(qian)(qian)预(xian)与异常(xian)模(kui)的(juan)线Q(ben)(yue)(huan)至?0%的(juan)(juan)(lu)(lu)(lu)(qian)(qian)(qian)分析(qi)(kui)研发(cai)(nin)l(zi)护(juan)成(qi)(mo)?/p> </p> <p>截止?.3.6.0版本Q?a href="//docs.yakult-sh.com.cn/advanced/TDgpt/" rel="nofollow noopener" target="_blank">TDgpt</a> 提(yun)(zong)Arima(mo)(mo)H(qiu)oltWinters(mo)(mo)LSTM(mo)(mo)MLP 以及Z(hong)Transformer架(duo)(nie)研(ye)的TDtsfm (TDengine time series foundation model) v1版和(nin)其他(zu)序(shi)(zhou)型Q(wu)及k-Sigma(mo)(mo)I(liang)nterquartile range(IQR)(mo)(mo)Grubbs(mo)(mo)SHESD(mo)(mo)Local Outlier Factor(LOF)(mo)(mo)Autoencoderq六U异(ning)检(xian)模(kui)。(yao)(ke)(xi)可以根?a href="//docs.yakult-sh.com.cn/advanced/TDgpt/" rel="nofollow noopener" target="_blank">TDgpt(liao)发指?/a>(nie)行接入(nie)研(ye)(xi)(yi)()他(yi)(fou)源的(zu)序(shi)(zhou)型(xi)算法?/p> </p> </blockquote> ]]></content:encoded> </item> <item> <title>Z TDgpt 时序数据体的电力需求预?/title> <link>//yakult-sh.com.cn/tdengine-engineering/28541.html</link> <dc:creator><![CDATA[derekchen]]></dc:creator> <pubDate>Tue, 25 Mar 2025 08:20:08 +0000</pubDate> <category><![CDATA[技术文?- 时序数据库]]></category> <category><![CDATA[tdgpt]]></category> <guid isPermaLink="false">//yakult-sh.com.cn/?p=28541</guid> <description><![CDATA[我们使用公开的UTSD数据集里面的电力需求数据,作ؓ预测法的数据来源,Z历史数据预测未来若干时的电力需求。]]></description> <content:encoded><![CDATA[<p><div id="ez-toc-container" class="ez-toc-v2_0_75 counter-hierarchy ez-toc-counter ez-toc-light-blue ez-toc-container-direction"> <div id="sqinis2ngw" class="ez-toc-title-container"> <p class="ez-toc-title" style="cursor:inherit">文章目录</p> <span class="ez-toc-title-toggle"><a href="#" class="ez-toc-pull-right ez-toc-btn ez-toc-btn-xs ez-toc-btn-default ez-toc-toggle" aria-label="Toggle Table of Content"><span class="ez-toc-js-icon-con"><span class=""><span class="eztoc-hide" style="display:none;">Toggle</span><span class="ez-toc-icon-toggle-span"><svg style="fill: #999;color:#999" xmlns="//www.w3.org/2000/svg" class="list-377408" width="20px" height="20px" viewBox="0 0 24 24" fill="none"><path d="M6 6H4v2h2V6zm14 0H8v2h12V6zM4 11h2v2H4v-2zm16 0H8v2h12v-2zM4 16h2v2H4v-2zm16 0H8v2h12v-2z" fill="currentColor"></path></svg><svg style="fill: #999;color:#999" class="arrow-unsorted-368013" xmlns="//www.w3.org/2000/svg" width="10px" height="10px" viewBox="0 0 24 24" version="1.2" baseProfile="tiny"><path d="M18.2 9.3l-6.2-6.3-6.2 6.3c-.2.2-.3.4-.3.7s.1.5.3.7c.2.2.4.3.7.3h11c.3 0 .5-.1.7-.3.2-.2.3-.5.3-.7s-.1-.5-.3-.7zM5.8 14.7l6.2 6.3 6.2-6.3c.2-.2.3-.5.3-.7s-.1-.5-.3-.7c-.2-.2-.4-.3-.7-.3h-11c-.3 0-.5.1-.7.3-.2.2-.3.5-.3.7s.1.5.3.7z"/></svg></span></span></span></a></span></div> <nav><ul class='ez-toc-list ez-toc-list-level-1 eztoc-toggle-hide-by-default' ><li class='ez-toc-page-1 ez-toc-heading-level-2'><a class="ez-toc-link ez-toc-heading-1" href="#" data-href="#Demo%E6%95%B0%E6%8D%AE%E9%9B%86%E5%87%86%E5%A4%87" >Demo数据集准?/a></li><li class='ez-toc-page-1 ez-toc-heading-level-2'><a class="ez-toc-link ez-toc-heading-2" href="#" data-href="#%E6%BC%94%E7%A4%BA%E7%8E%AF%E5%A2%83%E5%87%86%E5%A4%87" >演示环境准备</a><ul class='ez-toc-list-level-3' ><li class='ez-toc-heading-level-3'><a class="ez-toc-link ez-toc-heading-3" href="#" data-href="#%E7%8E%AF%E5%A2%83%E8%A6%81%E6%B1%82" >环境要求</a></li><li class='ez-toc-page-1 ez-toc-heading-level-3'><a class="ez-toc-link ez-toc-heading-4" href="#" data-href="#%E5%85%8B%E9%9A%86Demo%E4%BB%93%E5%BA%93%E5%88%B0%E6%9C%AC%E5%9C%B0" >克隆Demo仓库到本?/a></li><li class='ez-toc-page-1 ez-toc-heading-level-3'><a class="ez-toc-link ez-toc-heading-5" href="#" data-href="#%E8%BF%90%E8%A1%8C%E5%92%8C%E5%85%B3%E9%97%ADDemo" >q行和关闭Demo</a></li></ul></li><li class='ez-toc-page-1 ez-toc-heading-level-2'><a class="ez-toc-link ez-toc-heading-6" href="#" data-href="#%E8%BF%9B%E8%A1%8C%E6%BC%94%E7%A4%BA" >q行演示</a></li><li class='ez-toc-page-1 ez-toc-heading-level-2'><a class="ez-toc-link ez-toc-heading-7" href="#" data-href="#Demo%E8%84%9A%E6%9C%AC%E4%BD%BF%E7%94%A8%E8%AF%A6%E8%A7%A3" >Demo脚本使用详解</a><ul class='ez-toc-list-level-3' ><li class='ez-toc-heading-level-3'><a class="ez-toc-link ez-toc-heading-8" href="#" data-href="#%E8%84%9A%E6%9C%AC%E6%A6%82%E8%BF%B0" >脚本概述 </a></li><li class='ez-toc-page-1 ez-toc-heading-level-3'><a class="ez-toc-link ez-toc-heading-9" href="#" data-href="#%E5%8F%82%E6%95%B0%E8%AF%B4%E6%98%8E" >参数说明</a></li><li class='ez-toc-page-1 ez-toc-heading-level-3'><a class="ez-toc-link ez-toc-heading-10" href="#" data-href="#%E6%97%B6%E9%97%B4%E6%A0%BC%E5%BC%8F%E8%AF%B4%E6%98%8E" >旉格式说明 </a></li><li class='ez-toc-page-1 ez-toc-heading-level-3'><a class="ez-toc-link ez-toc-heading-11" href="#" data-href="#%E8%84%9A%E6%9C%AC%E6%89%A7%E8%A1%8C%E6%B5%81%E7%A8%8B" >脚本执行程</a></li></ul></li><li class='ez-toc-page-1 ez-toc-heading-level-2'><a class="ez-toc-link ez-toc-heading-12" href="#" data-href="#%E4%BD%BF%E7%94%A8%E6%9B%B4%E5%A4%9A%E7%9A%84%E6%95%B0%E6%8D%AE" >使用更多的数?/a></li><li class='ez-toc-page-1 ez-toc-heading-level-2'><a class="ez-toc-link ez-toc-heading-13" href="#" data-href="#%E7%BB%93%E8%AE%BA" >l论</a></li></ul></nav></div> </p> </p> <h2 class="gb-headline gb-headline-9e0fafd3 gb-headline-text">Demo数(ban)集(gong)?/h2> </p> <p>(xi)们(hua)(huan)(kua)公开?a href="//huggingface.co/datasets/thuml/UTSD" rel="nofollow noopener" target="_blank">UTSD数据?/a>里(yue)的电力(yun)求(yong)据,(huan)ؓ(fu)(chui)(juan)(geng)法的(juan)据来源,(nuo)Z历史(ji)数据(chui)(juan)未来若干(qi)的电力(yun)求。(yong)据集的采频ơؓ(fu)30(rou)钟Q(han)(huan)与旉(xi)x提供。ؓ(fu)了方便演C(hou)(juan)(ling)(chui)率?025-01-01 00:00:00开始向前倒(shu)生成旉(xi)I(chou)q存(jiang)储在TDengine(fen)应的表里?/p> </p> <p>数据集中包含5个文Ӟ我们使用~号最大的一个子集来完成演示。该数据文gQ放|于//github.com/taosdata/<a href="//yakult-sh.com.cn/tdgpt" data-internallinksmanager029f6b8e52c="16" title="TDgpt" target="_blank" rel="noopener">TDgpt</a>-demo仓库(fei)demo_data目录(juan)()请参考()(kou)的(qing)步骤导(jian)TDengine以完(yu)成演C(gan)。数据集(ju)(fei)统计信息如(juan)()</p> </p> <figure class="wp-block-image aligncenter size-large is-resized"><img loading="lazy" decoding="async" src="//yakult-sh.com.cn/wp-content/uploads/2025/03/image-62-1024x559.png" alt="Z TDgpt 时序数据体的电力需求预?- TDengine Database 时序数据? class="wp-image-28652" width="565" height="309" title="Z TDgpt 时序数据体的电力需求预?- TDengine Database 时序数据? srcset="//yakult-sh.com.cn/wp-content/uploads/2025/03/image-62-1024x559.png 1024w, //yakult-sh.com.cn/wp-content/uploads/2025/03/image-62-300x164.png 300w, //yakult-sh.com.cn/wp-content/uploads/2025/03/image-62-768x419.png 768w, //yakult-sh.com.cn/wp-content/uploads/2025/03/image-62.png 1272w" sizes="auto, (max-width: 565px) 100vw, 565px" /></figure> </p> </p> <h2 class="gb-headline gb-headline-a7ca32ad gb-headline-text">演示环境准备</h2> </p> <h3 class="gb-headline gb-headline-e5c9be59 gb-headline-text">环境要求</h3> </p> <p>您可(nuo)Z(hong)Linux(mo)Mac以及Windows操作(hu)pȝ(bai)完(shen)(yao)Demopȝ(bai)的运行。但(juan)Z(hong)(jiao)用docker-composeQ(shen)计算Z(hong)(an)(xia)安装(qi)(juan)(pin)软gQ?/p> </p> <ol class="wp-block-list" start="1"> <li>Git</li> </p> <li>Docker Engine: v20.10+</li> </p> <li>Docker Compose: v2.20+</li> </ol> </p> <p>Demo(juan)包?(juan)d(lie)ocker镜像 (TDengine, TDgpt, Grafana)Q以及一l(tong)于(tong)生预(xian)?异常(fu)(gui)(xian)结果的shell脚本(long)(mo)组件版(lian)的要求如下(xiao)Q?/p> </p> <figure class="wp-block-image aligncenter size-large"><img loading="lazy" decoding="async" width="1024" height="380" src="//yakult-sh.com.cn/wp-content/uploads/2025/03/image-63-1024x380.png" alt="Z TDgpt 时序数据体的电力需求预?- TDengine Database 时序数据? class="wp-image-28682" title="Z TDgpt 时序数据体的电力需求预?- TDengine Database 时序数据? srcset="//yakult-sh.com.cn/wp-content/uploads/2025/03/image-63-1024x380.png 1024w, //yakult-sh.com.cn/wp-content/uploads/2025/03/image-63-300x111.png 300w, //yakult-sh.com.cn/wp-content/uploads/2025/03/image-63-768x285.png 768w, //yakult-sh.com.cn/wp-content/uploads/2025/03/image-63-1536x570.png 1536w, //yakult-sh.com.cn/wp-content/uploads/2025/03/image-63.png 1664w" sizes="auto, (max-width: 1024px) 100vw, 1024px" /></figure> </p> </p> <h3 class="gb-headline gb-headline-60795402 gb-headline-text">克隆Demo仓(qin)(qi)(rou)本(she)?/h3> </p> <pre class="wp-block-code"><code>git clone //github.com/taosdata/TDgpt-demo cd TDgpt-demo chmod 775 analyse.sh</code></pre> </p> <p>文g夹下(xiao)包(yi)docker-compose.yml、tdengine.yml两个(fu)yml文g。docker-compose.yml 包(yi)了(ji)有一键启动demo(jia)(an)的镜像(fan)|(xiang)信(zhen),其引()tdengine.yml(huan)ؓ(fu)Grafana的数(kai)(rong)源配(qu)?/p> </p> <p>TDgpt-demo/demo_data下(pin)含三个csv文g(huan)Q(ben)electricity_demand.csv(mo)wind_power.csv(mo)ec2_failure.csvQ(ben),(xi)及三(bian)(fu)同前~(ke)sql脚本Q(ben)分(ke)(rou)对应电力需(ying)(cha)预(xian)、风力发(fa)()(dian)(xian)(pin)q维(huai)监(jia)异常(xian)(pin)(man)(xiu)?/p> </p> <p>TDgpt-demo/demo_dashboard(juan)(juan)包含了(jian)(juan)(juan)(bian)(bian)(fu)json文(yuan)g(huan)Qe(fen)lectricity_demand_forecast.json(mo)wind_power_forecast.json(mo)ec2_failure_anomaly.jsonQ,别对应(juan)(juan)(bian)(bian)(fu)场景的看ѝ?/p> </p> <p>docker-compose.yml中已l定(lei)(cen)了TDengine(yue)(gui)(sa)的(juan)(cen)化Ptdengine-dataQ待(yue)(gui)(sa)启动(nian)后,使用docker cp命odemo_data臛_器内使用(mo)?/p> </p> <h3 class="gb-headline gb-headline-117f8a73 gb-headline-text">q行和关闭Demo</h3> </p> <p><strong>(ping)意Q(ben)(ben)在q行demo前,h据您宿主机的架(duo)Q(ben)(ben)CPUcdQ(ben)(ben),~辑docker-compose.yml文gQ(ben)(ben)ؓ(fu)TDengine指定对应的platform(hong)(yong)(yong)Q(ben)(ben)linux/amd64Q(ben)(ben)Intel/AMD CPUQ(ben)(ben)或(shou)linux/arm64Q(ben)(ben)ARM CPUQ(ben)(ben)。TDgpt(jian)(xu)l一使用linux/amd64(hong)(yong)(yong)?/strong></p> </p> <p>q(ma)入docker-compose.yml文g(huan)所在的目录执行如下(duan)oQ启动TDengine、TDgpt和G(lan)rafana一体(yi)演(qi)环境Q?/p> </p> <pre class="wp-block-code"><code>docker-compose up -d</code></pre> </p> <p>首次q(ma)行Ӟ(e){待10s(yang)请执行下命oTDgpt的Anode节点(ping)册(xi)到T(bie)DengineQ?/p> </p> <pre class="wp-block-code"><code>docker exec -it tdengine taos -s "create anode 'tdgpt:6090'"</code></pre> </p> <p>在宿(juan)(juan)L(man)执行(juan)(juan)(pin)(duan)oQ初(pu)(pin)(huan)验试(dui)境的数据:</p> </p> <pre class="wp-block-code"><code>docker cp analyse.sh tdengine:/var/lib/taos docker cp demo_data tdengine:/var/lib/taos docker exec -it tdengine taos -s "source /var/lib/taos/demo_data/init_electricity_demand.sql"</code></pre> </p> <p>关闭演示环境Q请使(kua)(ke)Q?/p> </p> <pre class="wp-block-code"><code>docker-compose down</code></pre> </p> <h2 class="gb-headline gb-headline-518d22b4 gb-headline-text">q行演示</h2> </p> <ol class="wp-block-list" start="1"> <li>打(qin)览器,(cha)(qin)//localhost:3000Q(ben)(han)ƈ用默认的用户(jiong)名口?admin/admin dGrafana?/li> </p> <li>d成功后,q入路径”Home ?Dashboards”面Qƈ且导入electricity_demand_forecast.json文g?/li> </ol> </p> <figure class="wp-block-image aligncenter size-large is-resized"><img loading="lazy" decoding="async" src="//yakult-sh.com.cn/wp-content/uploads/2025/03/image-25-1024x344.png" alt="Z TDgpt 时序数据体的电力需求预?- TDengine Database 时序数据? class="wp-image-28543" width="768" height="258" title="Z TDgpt 时序数据体的电力需求预?- TDengine Database 时序数据? srcset="//yakult-sh.com.cn/wp-content/uploads/2025/03/image-25-1024x344.png 1024w, //yakult-sh.com.cn/wp-content/uploads/2025/03/image-25-300x101.png 300w, //yakult-sh.com.cn/wp-content/uploads/2025/03/image-25-768x258.png 768w, //yakult-sh.com.cn/wp-content/uploads/2025/03/image-25-1536x515.png 1536w, //yakult-sh.com.cn/wp-content/uploads/2025/03/image-25-2048x687.png 2048w" sizes="auto, (max-width: 768px) 100vw, 768px" /></figure> </p> </p> <ol class="wp-block-list" start="3"> <li>导入后,选择“electricity_demand”这个面ѝ面板已l配|好了真实倹{TDtsfm_1以及HoltWinters的预结果。当前只有真实值的数据曲线?/li> </ol> </p> </p> <ol class="wp-block-list" start="4"> <li>我们以analyze.sh脚本Q来q行预测。首先完成TDtsfm_1法的演C:</li> </ol> </p> <pre class="wp-block-code"><code>docker exec -it tdengine /var/lib/taos/analyse.sh --type forecast --db tdgpt_demo --table electricity_demand --stable single_val --algorithm tdtsfm_1 --params "fc_rows=48,wncheck=0" --start "2024-01-01" --window 30d --step 1d </code></pre> </p> <p>(juan)(juan)(juan)述shell脚本Q(han)(xiao)从指定的(qing)起始旉开始()(zhi)2024-01-01Q以(ji)一(cun)(juan)(juan)(juan)月(fei)数据ؓ输入(ai)Q()TDtsfm_1法(chui)测(gai)当前(shan)(shan)(juan)(juan)(juan)(yu)(cun)(chan)的(qing)?0mins(fei)电力需求()(zhi)p48(juan)(juan)(juan)数据点Q,直到(lang)辑ֈ(lang)electricity_demand (chen)中(fu)(lian)(yang)一(cun)(chan)的(qing)记录Q(han)ƈ结果写入electricity_demand_tdtsfm_1_result (chen)中(fu)(mo)执(qiao)(qiao)(chen)新(zhong)(fei)预前(shan)(shan)Q()(lian)会(xi)新徏/清(ping)对应(fei)结果表(mo)执(qiao)(qiao)(chen)()(beng)E(yu)(fu)持l(zi)在控制(bu)(hong)CQ按照天(juan)(juan)(juan)(hong)位推(sha)q(ma)输(wei)出(hong)(juan)(juan)(juan)的(qing)(jia)行l(zi)果Q?/p> </p> <pre class="wp-block-code"><code>taos> INSERT INTO tdgpt_demo.electricity_demand_tdtsfm_1_result SELECT _frowts, forecast(val, 'algorithm=tdtsfm_1,fc_rows=48,wncheck=0') FROM tdgpt_demo.electricity_deman WHERE ts >= '2024-01-12 00:00:00' AND ts < '2024-02-11 00:00:00' Insert OK, 48 row(s) affected (0.238208s)</code></pre> </p> <ol class="wp-block-list" start="5"> <li>Grafana的看板上Q配|刷新频率ؓ5sQ将动态显C预结果的黄色曲线Q直观呈C实际值的Ҏ。ؓ了展C清晎ͼh住command键点d下角的Real以及TDtsfm_1图例QMac下,Windows下请使用win键)Q从而只保留q两条曲U展C?/li> </ol> </p> <figure class="wp-block-image aligncenter size-large is-resized"><img loading="lazy" decoding="async" src="//yakult-sh.com.cn/wp-content/uploads/2025/03/image-28-1024x520.png" alt="Z TDgpt 时序数据体的电力需求预?- TDengine Database 时序数据? class="wp-image-28546" width="768" height="390" title="Z TDgpt 时序数据体的电力需求预?- TDengine Database 时序数据? srcset="//yakult-sh.com.cn/wp-content/uploads/2025/03/image-28-1024x520.png 1024w, //yakult-sh.com.cn/wp-content/uploads/2025/03/image-28-300x152.png 300w, //yakult-sh.com.cn/wp-content/uploads/2025/03/image-28-768x390.png 768w, //yakult-sh.com.cn/wp-content/uploads/2025/03/image-28-1536x780.png 1536w, //yakult-sh.com.cn/wp-content/uploads/2025/03/image-28-2048x1040.png 2048w" sizes="auto, (max-width: 768px) 100vw, 768px" /></figure> </p> </p> <figure class="wp-block-image aligncenter size-large is-resized"><img loading="lazy" decoding="async" src="//yakult-sh.com.cn/wp-content/uploads/2025/03/image-29-1024x523.png" alt="Z TDgpt 时序数据体的电力需求预?- TDengine Database 时序数据? class="wp-image-28547" width="768" height="392" title="Z TDgpt 时序数据体的电力需求预?- TDengine Database 时序数据? srcset="//yakult-sh.com.cn/wp-content/uploads/2025/03/image-29-1024x523.png 1024w, //yakult-sh.com.cn/wp-content/uploads/2025/03/image-29-300x153.png 300w, //yakult-sh.com.cn/wp-content/uploads/2025/03/image-29-768x392.png 768w, //yakult-sh.com.cn/wp-content/uploads/2025/03/image-29-1536x784.png 1536w, //yakult-sh.com.cn/wp-content/uploads/2025/03/image-29-2048x1045.png 2048w" sizes="auto, (max-width: 768px) 100vw, 768px" /></figure> </p> </p> <figure class="wp-block-image aligncenter size-large is-resized"><img loading="lazy" decoding="async" src="//yakult-sh.com.cn/wp-content/uploads/2025/03/image-30-1024x527.png" alt="Z TDgpt 时序数据体的电力需求预?- TDengine Database 时序数据? class="wp-image-28548" width="768" height="395" title="Z TDgpt 时序数据体的电力需求预?- TDengine Database 时序数据? srcset="//yakult-sh.com.cn/wp-content/uploads/2025/03/image-30-1024x527.png 1024w, //yakult-sh.com.cn/wp-content/uploads/2025/03/image-30-300x154.png 300w, //yakult-sh.com.cn/wp-content/uploads/2025/03/image-30-768x395.png 768w, //yakult-sh.com.cn/wp-content/uploads/2025/03/image-30-1536x791.png 1536w, //yakult-sh.com.cn/wp-content/uploads/2025/03/image-30-2048x1054.png 2048w" sizes="auto, (max-width: 768px) 100vw, 768px" /></figure> </p> </p> <figure class="wp-block-image aligncenter size-large is-resized"><img loading="lazy" decoding="async" src="//yakult-sh.com.cn/wp-content/uploads/2025/03/image-31-1024x524.png" alt="Z TDgpt 时序数据体的电力需求预?- TDengine Database 时序数据? class="wp-image-28549" width="768" height="393" title="Z TDgpt 时序数据体的电力需求预?- TDengine Database 时序数据? srcset="//yakult-sh.com.cn/wp-content/uploads/2025/03/image-31-1024x524.png 1024w, //yakult-sh.com.cn/wp-content/uploads/2025/03/image-31-300x154.png 300w, //yakult-sh.com.cn/wp-content/uploads/2025/03/image-31-768x393.png 768w, //yakult-sh.com.cn/wp-content/uploads/2025/03/image-31-1536x786.png 1536w, //yakult-sh.com.cn/wp-content/uploads/2025/03/image-31-2048x1048.png 2048w" sizes="auto, (max-width: 768px) 100vw, 768px" /></figure> </p> </p> <ol class="wp-block-list" start="6"> <li>完成HoltWinters模型的演C:</li> </ol> </p> <pre class="wp-block-code"><code>docker exec -it tdengine /var/lib/taos/analyse.sh --type forecast --db tdgpt_demo --table electricity_demand --stable single_val --algorithm holtwinters --params "rows=48,period=48,wncheck=0,trend=add,seasonal=add" --start "2024-01-01" --window 30d --step 1d </code></pre> </p> <p>(juan)第四步cMQ(ben)HoltWinters模型(hao)(gong)(nian)(hao)输出(hong)(xian)(xu)(bai)果ƈ(gou)呈(bao)(bang)(she)看(lian)板上()(mo)从预测l(zi)果(juan)可(xi)看(lian)刎(bang)(ren)TDtsfm_1Ҏ(gui)(rong)的预测_(qi)ֺ(hai)显著(xian)优于传统(bai)的统(bai)计学(kou)Ҏ(gui)(chen)(chen)HoltWinters(mo)除(bang)预(xian)(xu)(qian)外Q(ben)HoltWinters法(chen)(chen)的最大问题(kui)需要非(ning)精l(zi)(gong)的对(hong)数q行调(xie)评估(ji)Q(ben)否则还Ҏ(gui)出(hong)(bang)(juan)(pin)(juan)这U(zui)J(huo)(fa)生的预测值奇异点(mo)?/p> </p> <figure class="wp-block-image aligncenter size-large is-resized"><img loading="lazy" decoding="async" src="//yakult-sh.com.cn/wp-content/uploads/2025/03/image-26-1024x526.png" alt="Z TDgpt 时序数据体的电力需求预?- TDengine Database 时序数据? class="wp-image-28544" width="768" height="395" title="Z TDgpt 时序数据体的电力需求预?- TDengine Database 时序数据? srcset="//yakult-sh.com.cn/wp-content/uploads/2025/03/image-26-1024x526.png 1024w, //yakult-sh.com.cn/wp-content/uploads/2025/03/image-26-300x154.png 300w, //yakult-sh.com.cn/wp-content/uploads/2025/03/image-26-768x395.png 768w, //yakult-sh.com.cn/wp-content/uploads/2025/03/image-26-1536x790.png 1536w, //yakult-sh.com.cn/wp-content/uploads/2025/03/image-26-2048x1053.png 2048w" sizes="auto, (max-width: 768px) 100vw, 768px" /></figure> </p> </p> <p>Z鼠(chuo)圈选(ling)方式Q(shen)们可以查(lu)(yu)D|间(cun)的细_度(hai)预(juan)l果ҎQ?/p> </p> <figure class="wp-block-image aligncenter size-large is-resized"><img loading="lazy" decoding="async" src="//yakult-sh.com.cn/wp-content/uploads/2025/03/image-27-1024x521.png" alt="Z TDgpt 时序数据体的电力需求预?- TDengine Database 时序数据? class="wp-image-28545" width="768" height="391" title="Z TDgpt 时序数据体的电力需求预?- TDengine Database 时序数据? srcset="//yakult-sh.com.cn/wp-content/uploads/2025/03/image-27-1024x521.png 1024w, //yakult-sh.com.cn/wp-content/uploads/2025/03/image-27-300x153.png 300w, //yakult-sh.com.cn/wp-content/uploads/2025/03/image-27-768x391.png 768w, //yakult-sh.com.cn/wp-content/uploads/2025/03/image-27-1536x781.png 1536w, //yakult-sh.com.cn/wp-content/uploads/2025/03/image-27-2048x1041.png 2048w" sizes="auto, (max-width: 768px) 100vw, 768px" /></figure> </p> </p> <p>您也(mei)可以试(cu)其(duo)(qu)(he)(chen)或(ci)(kui),来找到最(ying)合适自己场(man)(xiu)的(he)(chen)(xia)(shen)(kui)?/p> </p> <h2 class="gb-headline gb-headline-a5e518eb gb-headline-text">Demo脚本使用详解</h2> </p> <h3 class="gb-headline gb-headline-827fcf62 gb-headline-text">脚本概述 </h3> </p> <p>analyse.sh脚本(long)用于(she)?TDengine 数据(qian)上(jia)行(zu)(qian)列预测和异(ning)检分析(xie)(ren)支持滑动H(yu)(he)处(tong)。(bei)要功能(jie)(ma):</p> </p> <ul class="wp-block-list"> <li>旉序(gan)预 Q(qing)(jiao)?HoltWinters {算法(die)(yu)行(shen)来值预??/li> </p> <li>异常?Q(ben)?k-Sigma {算法识别数(kai)(rong)异常点(yan) (mo)?/li> </p> <li>自动H口滑动 Q支持自定义H口大小和步长进行连l分析?/li> </ul> </p> <h3 class="gb-headline gb-headline-d437b0a2 gb-headline-text">参数说明</h3> </p> <figure class="wp-block-image aligncenter size-large is-resized"><img loading="lazy" decoding="async" src="//yakult-sh.com.cn/wp-content/uploads/2025/03/image-32-1024x548.png" alt="Z TDgpt 时序数据体的电力需求预?- TDengine Database 时序数据? class="wp-image-28550" width="768" height="411" title="Z TDgpt 时序数据体的电力需求预?- TDengine Database 时序数据? srcset="//yakult-sh.com.cn/wp-content/uploads/2025/03/image-32-1024x548.png 1024w, //yakult-sh.com.cn/wp-content/uploads/2025/03/image-32-300x161.png 300w, //yakult-sh.com.cn/wp-content/uploads/2025/03/image-32-768x411.png 768w, //yakult-sh.com.cn/wp-content/uploads/2025/03/image-32-1536x822.png 1536w, //yakult-sh.com.cn/wp-content/uploads/2025/03/image-32.png 1924w" sizes="auto, (max-width: 768px) 100vw, 768px" /></figure> </p> </p> <p>TDengine推荐使用(zan)(ping)(hai)表来q行(suo)(suo)据建模。(xiao)()此,Demo(juan)(juan)徏(huan)立(yu)(juan)(juan)(juan)(juan)名(xi)(juan)(juan)?single_val 的超(he)U表Q包含ts (timestampc(ling)(ling)d(hui)) 和val (floatc(ling)(ling)d(hui))Q以及标{(jiang)֮(cen)scene (varchar (64) )。现(qu)段TDgpt只支(xia)持单列D(wei)(wei)入输(wei)(wei)出,因此q个(zan)(ping)(hai)表可以作(juan)(juan)所有源(suo)(suo)据表和l(zi)果表的l(zi)构定义。(xiao)(dao)表的表名(xi)(juan)(juan)tag名称保持(juan)(juan)(nie)即可?/p> </p> <p>db参数指定了源(zan)(suo)据(chen)和l(zi)果(yong)(chen)隶(ba)的(suo)据库(tuo)结果(he)以(mo)源(zan)(chen)名U】_(mo)算法名U】_(mo)result(mo)(jia)式存(xian)。Grafana(yue)面(kui)通过(mo)询l(zi)果(yong)(chen)实(lin)现分析结果和原始(suo)据(fei)(fan)比?/p> </p> <p>(juan)(lei)情况下Q(ben)(han)于非(jian)填,用户在demoq程(juan)只(an)要设|(xiang)?#8211;start(hong)(yong)以节(lu)运(xiu)行时(xiang)(cui)对于必(sui)(sui)项Q(ben)()(hong)考示例D行()|(xiang)?/p> </p> <h3 class="gb-headline gb-headline-ff6343b8 gb-headline-text">旉格式说明 </h3> </p> <p>step和window参(yong)(yong)(kai)指(lei)的滑动步长(kua)分析H(yu)(yi)大小需W(heng)合(hao)如下参(yong)(yong)(kai)U定(lei)Q?/p> </p> <figure class="wp-block-image aligncenter size-large is-resized"><img loading="lazy" decoding="async" src="//yakult-sh.com.cn/wp-content/uploads/2025/03/image-33-1024x204.png" alt="Z TDgpt 时序数据体的电力需求预?- TDengine Database 时序数据? class="wp-image-28551" width="768" height="153" title="Z TDgpt 时序数据体的电力需求预?- TDengine Database 时序数据? srcset="//yakult-sh.com.cn/wp-content/uploads/2025/03/image-33-1024x204.png 1024w, //yakult-sh.com.cn/wp-content/uploads/2025/03/image-33-300x60.png 300w, //yakult-sh.com.cn/wp-content/uploads/2025/03/image-33-768x153.png 768w, //yakult-sh.com.cn/wp-content/uploads/2025/03/image-33-1536x306.png 1536w, //yakult-sh.com.cn/wp-content/uploads/2025/03/image-33.png 1910w" sizes="auto, (max-width: 768px) 100vw, 768px" /></figure> </p> </p> <h3 class="gb-headline gb-headline-02f208e6 gb-headline-text">脚本执行程</h3> </p> <pre class="wp-block-code"><code>graph TDgpt_Demo A[开始] --> B[参数解析与验证] B --> C{是否指定start?} C -->|否| D[查询最时间戳] C -->|是| E[转换旉格式] D --> E E --> F[计算旉H口] F --> G[生成l果表] G --> H{是否到达数据l点?} H -->|否| I[生成q执行SQL] I --> H H -->|是| J[输出完成信息]</code></pre> </p> <h2 class="gb-headline gb-headline-7807291e gb-headline-text">使用(xuan)多的数(kai)?/h2> </p> <p>(hong)考「运行和(wei)闭Demo」(qu)节里(xin)electricity_demand.sql脚本的内(yue),保按照规定(luo)式数(rong)准备ؓ(fu)csv(luo)式Q(ben)(ben)逗号分隔Q(ben)(ben)值需(ying)要(quan)英(hong)引(hong)hhQ(ben)(ben),卛_(chong)数(rong)导(wei)TDengine。然(yang),(xuan)用「进行演C」(qu)节中的方法(chen)生成预测l果Q(ben)(ben)ƈ(gou)调整Grafana中的看板以实现和(yue)际(suo)据的对比?/p> </p> <h2 class="gb-headline gb-headline-859e10e6 gb-headline-text">l论</h2> </p> <p>在本(long)(kou)中Q(shen)(duo)们展C(gan)Z(hong)(huan)(huan)用TDgpt来进行电(e)需求预(xian)的完(shen)(qi)(xian)程(mo)从(juan)可(chi)以看刎ͼZ(hong)TDgpt 来构建时(qian)数据分析,能够以SQL(kou)式实现(juan)应用的便捷集成Q还(hong)以用Grafana q行展示Q(han)大降(huan)(huan)开(hong)(cai)(qian)用时序数据预测和(han)(shu)常检(xian)的(xi)(qi)(long)(mo)?/p> </p> <p>从预(xian)效果来(sa)看,Ztransformer架构的(dan)(cong)练(zhou)型TDtsfm_1在()的数集(ru)展示出显(jue)优(bang)Holtwinters(zhou)型的(juan)果。在不同(qian)的实际(yi)景下Q用户需要针(kuo)Ҏ据特点,针(guo)(zhou)型(geng)法q行(kun)择和(han)数调优,也可以选择不同(qian)的算法或(shou)(zhou)型q行试?/p> </p> <p>TDengine 的企业版中,TDgpt (hao)ؓ用户提供(xuan)多的选(qun)Q?/p> </p> <ol class="wp-block-list" start="1"> <li>模型(kun)(qun)(han)(han)(zhao)。模型选(qun)(han)(han)(zhao)可(chi)(chi)以自(e)根(rong)用L(zha)史数(ban)(ban),对购(cen)的所(lian)(qun)型进行准性评估。用户可(chi)(chi)(kun)(qun)(han)(han)(lian)(kun)合自己场景(ou)的模型或法q(ma)行部v(jiang)和应用?/li> </p> <li>TDtsfm_1(nie)研模型(fei)重(cong)练(hong)微调。TDtsfm_1(nuo)Z(xian)量()(zu)序数(ban)(kan)q(du)(bang)(bang)(bang)预(cong)练Q在(chan)部(rou)(gong)(man)景下(sui)比(bang)(bang)(bang)传l(zi)的(lian)(lian)器(sa)学习和统(cong)预(xian)()型都(shi)会有显著(xian)(fei)准(yun)率优势。如果用(ke)(ke)户对(bang)(bang)(bang)模型预(xian)准(yun)度(lian)(lian)(qun)高的(xia)求Q可以申误(cen)?TDgpt 企业(gu)版的(chui)训l(zi)服(qi)务。(jiao)()用(ke)(ke)L(feng)(she)景历(gong)数(ban)(kan)q(du)(chui)训l(zi),(she)特定场(man)景下(fei)预(xian)()果可(jian)更佟?/li> </p> <li>W三方解x案。涛思数据联合国内外时序分析/异常专业厂家、研I机构,为用h供专业的分析解决ҎQ包括落地过E中的实施服务等?/li> </ol> </p> <p>关于(lu)企业版更多信息,点(juan)()(nan)按钮Q咨询解(cong)x(chong)(nan)案专家?/p> </p> <p><div id="sqinis2ngw" class="gb-container gb-container-41506e9a advice"><a class="gb-container-link" rel="nofollow"></a></p> <div id="sqinis2ngw" class="gb-container gb-container-5d703363"> <p><span class="gb-button gb-button-12241d8d gb-button-text">(ku)即咨询</span></p> </div> <p> </div> </p> </p> <blockquote class="wp-block-quote is-layout-flow wp-block-quote-is-layout-flow"> <p><strong>关于背景</strong></p> </p> <p>电(dian)需求预作(hu)(juan)(hong)(xi)能源(yao)理的(luo)心工具Q(ben)(han)(luo)心(xi)DI电(shu)(shu)力系l的全生命周(lian)。在资(juan)(zan)配置层面Q(ben)通过(beng)(beng)_և(chui)判用电(shu)(shu)势Q(ben)(han)(chi)(mei)化发电(shu)(shu)(cong)施布(fu)(long)(juan)(tiao)(shu)(shu)|升U(wan)节奏(xi)避免前投资造成的资源闲|或(biao)后(wu)引发的供(zong)应缺口,典型场景(juan)可(chi)使基(yun)(cong)施投资效率提(su)15%-20%(mo)对于(tiao)(shu)(shu)力运(nie)商而言Q(ben)负(zhai)荷预支撑着(xi)(tiao)料(dang)(yong)购到(lang)(lian)(hong)调(fu)的动态优化,在火力发电领域已实现吨煤发电(shu)(shu)?%(xi)上的能效提升,同(shen)通(beng)(beng)削峰填谷(yao)低늽备用定w需求,显著(xian)压羃pȝq(song)成本(mo)?nbsp;</p> </p> <p>(she)能(shi)(jing)(su)全维度,预(juan)(ge)(lian)构(ying)v(dao)()(())(dian)供(yun)的(tong)冲机制。短(lian)预(xian)(nai)差每降低1个百分点(yan)Q(ben)(han)应减的(qing)紧急调峰(ban)(lian)可(cha)֌(kui)(nuo)电|日(yuan)均运(xiu)(nie)费()(())的(qing)3%-5%Q(ben)这(she)应Ҏ端天气或H(huo)(bang)g(huan)(zu)(duo)为关键。而中(fu)(xiang)期预(juan)则ؓ跨区(kui)(nuo)电(e)(shui)(xian)、(huo)能设施(jie)|提供决(yu){基U,(lian)效~解l(zi)构(ying)(hao)缺()(())风险。市(she)环(yi)境()(fu)Q(ben)预(xian)(nai)(shi)(e)直接{(jiang)(chen)ؓl(zi)济收益Q(ben)(han)()(())企业通(yun)(zu)前96(zu)段(lu)荷预(juan)(mei)化(ge)h(huan){(jian)Q(ben)(han)()()(())(dian)(dui)货(chuai)市场中可额外获(feng)10%-18%的h(huan)格套利空_工商()(())户则借助(lu)荷Ҏ分(man)(su)(yue)用能方(jin),(yue)(tong)q(qian)(cun)()(())费支出5%-10%的降q(qian)?/p> </p> <p>本文提(yun)(zong)Z docker-compose (jian)(jian)速部|?TDgp 体验(bi)试环境的(juan)引,q基(huang)于这个环(li)和真实(lin)(juan)据,展示日前预(juan)电(dian)需(cha)的全过E()(zhu)于(chan)家(jian)(jian)速掌?TDgptQ迅速让自己(fan)拥有AI驱动的(juan)(lu)(qian)数据预(yu)(ya)异常的能力?/p> </p> <p><strong>关于TDgpt</strong></p> </p> <p>TDgpt ?TDengine (cong)置(fei)(fei)(juan)(lu)(qian)(qi)(rong)分(ke)析智能体(hong)Q(ben)(ben)(han)Z(lu) TDengine (fei)(fei)(juan)(lu)(qian)(qi)(rong)查询功能,(ren)通(yun)(yun) SQL 提(yun)q行时可动态扩展(li)(nin)切(pian)(fei)(fei)(juan)(lu)(qian)(qi)(rong)高U分(ke)析(yao)(qing)能力Q(ben)(ben)(han)括时(lu)(qian)(qi)(rong)预(pin)(nin)时序数(ban)(kan)异(xiao)。通(yun)(yun)预置(fei)(fei)(juan)(lu)(qian)大(zhou)型(mo)大语言(zhou)型(mo)机器学(cen)、传l(xi)(qing)法Q(ben)(ben)TDgpt 能帮助工(fa)E(pin)(she)?0分(qian)(cong)完(xi)时(lu)(qian)预与(ya)异(xiao)模型的(qing)上线Q(ben)(ben)降低至?0%(fei)(fei)(juan)(lu)(qian)分(ke)析模型研(hong)(cai)(nin)l护(xi)本(long)(mo)?/p> </p> <p>截止?.3.6.0版本Q?a href="//docs.yakult-sh.com.cn/advanced/TDgpt/" rel="nofollow noopener" target="_blank">TDgpt</a> 提供Arima(mo)(mo)H(qiu)oltWinters(mo)(mo)LSTM(mo)(mo)MLP (xi)及(qiang)(nuo)ZTransformer架构(nie)研的TDtsfm (TDengine time series foundation model) v1(jian)和其他(qu)(zu)序(shi)(zhou)型Q以及k(hong)-Sigma(mo)(mo)Interquartile range(IQR)(mo)(mo)Grubbs(mo)(mo)S(jing)HESD(mo)(mo)Local Outlier Factor(LOF)(mo)(mo)Autoencoderq六U异常检模型。用户可(xi)根(rong)?a href="//docs.yakult-sh.com.cn/advanced/TDgpt/" rel="nofollow noopener" target="_blank">TDgpt开发指(ti)?/a>(nie)行()入(ai)(nie)研或其(xi)开源的(qing)(zu)序(shi)(zhou)型或(jian)(ping)?/p> </p> </blockquote> ]]></content:encoded> </item> </channel> </rss>