TDgpt预测能否相信和可用

[TDuser_fAGX_8210 (李勇刚)]

taos3.3.7.0新建1点表，其结构如下：

show create table yc_d1150000820_z_151yc \G;

*************************** 1.row ***************************
Table: yc_d1150000820_z_151yc
Create Table: CREATE TABLE yc_d1150000820_z_151yc USING analog (dev, bay, station, terno) TAGS (“D1150000820”, “bay1050000034”, “sub1080000005”, 153)
Query OK, 1 row(s) in set (0.000644s)

taos> desc yc_d1150000820_z_151yc;

field | type | length | note | encode | compress | level |

ts | TIMESTAMP | 8 | | delta-i | lz4 | medium |
data | FLOAT | 4 | | delta-d | lz4 | medium |
quality | INT | 4 | | simple8b | lz4 | medium |
dev | VARCHAR | 48 | TAG | disabled | disabled | disabled |
bay | VARCHAR | 48 | TAG | disabled | disabled | disabled |
station | VARCHAR | 48 | TAG | disabled | disabled | disabled |
terno | INT | 4 | TAG | disabled | disabled | disabled |
Query OK, 7 row(s) in set (0.000948s)

该点每3秒插入一条数据，数据递增，从0-1000，形成锯齿波，分别用chronos、tdtsfm_1、timemoe-fc三种算法去预测，预测结果如下：

taos> select _frowts, forecast(data,“algo=chronos,rows=10”) from yc_d1150000820_z_151yc;

_frowts | forecast(data,“algo=chronos,rows=10”) |

2025-09-03 14:46:01.814 | 884.632446289062 |
2025-09-03 14:46:04.896 | 886.880249023438 |
2025-09-03 14:46:07.978 | 889.1279296875 |
2025-09-03 14:46:11.060 | 889.1279296875 |
2025-09-03 14:46:14.142 | 891.375732421875 |
2025-09-03 14:46:17.224 | 891.375732421875 |
2025-09-03 14:46:20.306 | 891.375732421875 |
2025-09-03 14:46:23.388 | 893.623413085938 |
2025-09-03 14:46:26.470 | 893.623413085938 |
2025-09-03 14:46:29.552 | 895.871215820312 |
Query OK, 10 row(s) in set (0.359570s)

taos> select _frowts, forecast(data,“algo=tdtsfm_1,rows=10”) from yc_d1150000820_z_151yc;

_frowts | forecast(data,“algo=tdtsfm_1,rows=10”) |

2025-09-03 14:46:01.814 | 953.45703125 |
2025-09-03 14:46:04.896 | 947.710327148438 |
2025-09-03 14:46:07.978 | 948.882690429688 |
2025-09-03 14:46:11.060 | 950.314208984375 |
2025-09-03 14:46:14.142 | 946.8359375 |
2025-09-03 14:46:17.224 | 943.424926757812 |
2025-09-03 14:46:20.306 | 949.297485351562 |
2025-09-03 14:46:23.388 | 950.871826171875 |
2025-09-03 14:46:26.470 | 949.113952636719 |
2025-09-03 14:46:29.552 | 950.952026367188 |
Query OK, 10 row(s) in set (0.102288s)

taos> select _frowts, forecast(data,“algo=timemoe-fc,rows=10”) from yc_d1150000820_z_151yc;

_frowts | forecast(data,“algo=timemoe-fc,rows=10”) |

2025-09-03 14:46:01.814 | 54.5072021484375 |
2025-09-03 14:46:04.896 | 55.5149841308594 |
2025-09-03 14:46:07.978 | 56.2104797363281 |
2025-09-03 14:46:11.060 | 56.8955383300781 |
2025-09-03 14:46:14.142 | 59.9058532714844 |
2025-09-03 14:46:17.224 | 63.0441589355469 |
2025-09-03 14:46:20.306 | 62.9778747558594 |
2025-09-03 14:46:23.388 | 64.1473388671875 |
2025-09-03 14:46:26.470 | 60.9953002929688 |
2025-09-03 14:46:29.552 | 63.0239562988281 |
Query OK, 10 row(s) in set (4.787222s)

问题：1.预测不准；2.预测结果差异太大。请消缺或解释原因.

[TDuser_fAGX_8210 (李勇刚)]

原始数据见附件。

[Jade (Jade Chen)]

好的我们看下

[Haojun_Liao (廖浩均)]

尝试使用规模更大的 chronos 或 timemoe 的时序基础模型测试一下，可能结果会更好一些。方法请参见： 部署时序基础模型 | TDengine 文档 | 涛思数据
TDgpt 直接调用的开源时序基础模型。对于您提到的数据的预测的可信度问题，建议计算 MAPE,并给出量化的结果。
不同的时序模型由于其训练使用的时序数据不同，参数规模不同，可能表现不一样。需要结合具体的场景，选择最合适的时序模型。
如果所有所有的模型在您的业务数据集合上表现均不理想，可能需要考虑应用时序模型在您的业务数据集上进行 fine-tuning。不同的时序模型 fine-tuning方式有些许差别，需要参考其具体操作说明，例如针对 timemoe操作方式参见 GitHub - Time-MoE/Time-MoE: [ICLR 2025 Spotlight] Official implementation of "Time-MoE: Billion-Scale Time Series Foundation Models with Mixture of Experts" fine-tuning 部分。

[TDuser_fAGX_8210 (李勇刚)]

测试里包含了chronos和timemoe哟

[Haojun_Liao (廖浩均)]

明白，我的意思是您可以 考虑使用参数更大的时序模型。因为每个时序基础模型都有不同规模参数的版本。参数多一些，性能会好一些。TDgpt 设置的使用的默认版本一般都不是规模最大的版本。