ETDataset/ett|电力变压器数据集|时间序列预测数据集

---
annotations_creators:
- no-annotation
language_creators:
- found
language: []
license:
- cc-by-4.0
multilinguality:
- monolingual
pretty_name: Electricity Transformer Temperature
size_categories:
- 1K<n<10K
source_datasets:
- original
task_categories:
- time-series-forecasting
task_ids:
- univariate-time-series-forecasting
- multivariate-time-series-forecasting
dataset_info:
- config_name: h1
features:
- name: start
dtype: timestamp[s]
- name: target
sequence: float32
- name: feat_static_cat
sequence: uint64
- name: feat_dynamic_real
sequence:
sequence: float32
- name: item_id
dtype: string
splits:
- name: train
num_bytes: 241978
num_examples: 1
- name: test
num_bytes: 77508960
num_examples: 240
- name: validation
num_bytes: 33916080
num_examples: 120
download_size: 2589657
dataset_size: 111667018
- config_name: h2
features:
- name: start
dtype: timestamp[s]
- name: target
sequence: float32
- name: feat_static_cat
sequence: uint64
- name: feat_dynamic_real
sequence:
sequence: float32
- name: item_id
dtype: string
splits:
- name: train
num_bytes: 241978
num_examples: 1
- name: test
num_bytes: 77508960
num_examples: 240
- name: validation
num_bytes: 33916080
num_examples: 120
download_size: 2417960
dataset_size: 111667018
- config_name: m1
features:
- name: start
dtype: timestamp[s]
- name: target
sequence: float32
- name: feat_static_cat
sequence: uint64
- name: feat_dynamic_real
sequence:
sequence: float32
- name: item_id
dtype: string
splits:
- name: train
num_bytes: 967738
num_examples: 1
- name: test
num_bytes: 1239008640
num_examples: 960
- name: validation
num_bytes: 542089920
num_examples: 480
download_size: 10360719
dataset_size: 1782066298
- config_name: m2
features:
- name: start
dtype: timestamp[s]
- name: target
sequence: float32
- name: feat_static_cat
sequence: uint64
- name: feat_dynamic_real
sequence:
sequence: float32
- name: item_id
dtype: string
splits:
- name: train
num_bytes: 967738
num_examples: 1
- name: test
num_bytes: 1239008640
num_examples: 960
- name: validation
num_bytes: 542089920
num_examples: 480
download_size: 9677236
dataset_size: 1782066298
---

# Dataset Card for [Electricity Transformer Temperature](https://github.com/zhouhaoyi/ETDataset)

## Table of Contents
- [Table of Contents](#table-of-contents)
- [Dataset Description](#dataset-description)
- [Dataset Summary](#dataset-summary)
- [Supported Tasks and Leaderboards](#supported-tasks-and-leaderboards)
- [Languages](#languages)
- [Dataset Structure](#dataset-structure)
- [Data Instances](#data-instances)
- [Data Fields](#data-fields)
- [Data Splits](#data-splits)
- [Dataset Creation](#dataset-creation)
- [Curation Rationale](#curation-rationale)
- [Source Data](#source-data)
- [Annotations](#annotations)
- [Personal and Sensitive Information](#personal-and-sensitive-information)
- [Considerations for Using the Data](#considerations-for-using-the-data)
- [Social Impact of Dataset](#social-impact-of-dataset)
- [Discussion of Biases](#discussion-of-biases)
- [Other Known Limitations](#other-known-limitations)
- [Additional Information](#additional-information)
- [Dataset Curators](#dataset-curators)
- [Licensing Information](#licensing-information)
- [Citation Information](#citation-information)
- [Contributions](#contributions)

## Dataset Description

- **Homepage:** [Electricity Transformer Dataset](https://github.com/zhouhaoyi/ETDataset)
- **Repository:** https://github.com/zhouhaoyi/ETDataset
- **Paper:** [Informer: Beyond Efficient Transformer for Long Sequence Time-Series Forecasting](https://arxiv.org/abs/2012.07436)
- **Point of Contact:** [Haoyi Zhou](mailto:[e-mail])

### Dataset Summary

The electric power distribution problem is the distribution of electricity to different areas depending on its sequential usage. But predicting the future demand of a specific area is difficult, as it varies with weekdays, holidays, seasons, weather, temperatures, etc. However, no existing method can perform a long-term prediction based on super long-term real-world data with high precision. Any false predictions may damage the electrical transformer. So currently, without an efficient method to predict future electric usage, managers have to make decisions based on the empirical number, which is much higher than the real-world demands. It causes unnecessary waste of electric and equipment depreciation. On the other hand, the oil temperatures can reflect the condition of the Transformer. One of the most efficient strategies is to predict how the electrical transformers' oil temperature is safe and avoid unnecessary waste. As a result, to address this problem, the authors and Beijing Guowang Fuda Science & Technology Development Company have provided 2-years worth of data.

Specifically, the dataset combines short-term periodical patterns, long-term periodical patterns, long-term trends, and many irregular patterns. The dataset are obtained from 2 Electricity Transformers at 2 stations and come in an `1H` (hourly) or `15T` (15-minute) frequency containing 2 year * 365 days * 24 hours * (4 for 15T) times = 17,520 (70,080 for 15T) data points.

The target time series is the **O**il **T**emperature and the dataset comes with the following 6 covariates in the univariate setup:
* **H**igh **U**se**F**ul **L**oad
* **H**igh **U**se**L**ess **L**oad
* **M**iddle **U**se**F**ul **L**oad
* **M**iddle **U**se**L**ess **L**oad
* **L**ow **U**se**F**ul **L**oad
* **L**ow **U**se**L**ess **L**oad

### Dataset Usage

To load a particular variant of the dataset just specify its name e.g:

```python
load_dataset("ett", "m1", multivariate=False) # univariate 15-min frequency dataset from first transformer
```

or to specify a prediction length:

```python
load_dataset("ett", "h2", prediction_length=48) # multivariate dataset from second transformer with prediction length of 48 (hours)
```

### Supported Tasks and Leaderboards

The time series data is split into train/val/test set of 12/4/4 months respectively. Given the prediction length (default: 1 day (24 hours or 24*4 15T)) we create rolling windows of this size for the val/test sets.

#### `time-series-forecasting`

##### `univariate-time-series-forecasting`

The univariate time series forecasting tasks involves learning the future one dimensional `target` values of a time series in a dataset for some `prediction_length` time steps. The performance of the forecast models can then be validated via the ground truth in the `validation` split and tested via the `test` split. The covriates are stored in the `feat_dynamic_real` key of each time series.

##### `multivariate-time-series-forecasting`

The multivariate time series forecasting task involves learning the future vector of `target` values of a time series in a dataset for some `prediction_length` time steps. Similar to the univariate setting the performance of a multivariate model can be validated via the ground truth in the `validation` split and tested via the `test` split.

### Languages

## Dataset Structure

### Data Instances

A sample from the training set is provided below:

```python
{
'start': datetime.datetime(2012, 1, 1, 0, 0),
'target': [14.0, 18.0, 21.0, 20.0, 22.0, 20.0, ...],
'feat_static_cat': [0],
'feat_dynamic_real': [[0.3, 0.4], [0.1, 0.6], ...],
'item_id': 'OT'
}
```

### Data Fields

For the univariate regular time series each series has the following keys:

* `start`: a datetime of the first entry of each time series in the dataset
* `target`: an array[float32] of the actual target values
* `feat_static_cat`: an array[uint64] which contains a categorical identifier of each time series in the dataset
* `feat_dynamic_real`: optional array of covariate features
* `item_id`: a string identifier of each time series in a dataset for reference

For the multivariate time series the `target` is a vector of the multivariate dimension for each time point.

### Data Splits

The time series data is split into train/val/test set of 12/4/4 months respectively.

## Dataset Creation

### Curation Rationale

Develop time series methods that can perform a long-term prediction based on super long-term real-world data with high precision.

### Source Data

#### Initial Data Collection and Normalization

[More Information Needed]

#### Who are the source language producers?

[More Information Needed]

### Annotations

#### Annotation process

[More Information Needed]

#### Who are the annotators?

[More Information Needed]

### Personal and Sensitive Information

[More Information Needed]

## Considerations for Using the Data

### Social Impact of Dataset

[More Information Needed]

### Discussion of Biases

[More Information Needed]

### Other Known Limitations

[More Information Needed]

## Additional Information

### Dataset Curators

* [Haoyi Zhou](mailto:[e-mail])

### Licensing Information

[Creative Commons Attribution 4.0 International](https://creativecommons.org/licenses/by/4.0/legalcode)

### Citation Information

```tex
@inproceedings{haoyietal-informer-2021,
author = {Haoyi Zhou and
Shanghang Zhang and
Jieqi Peng and
Shuai Zhang and
Jianxin Li and
Hui Xiong and
Wancai Zhang},
title = {Informer: Beyond Efficient Transformer for Long Sequence Time-Series Forecasting},
booktitle = {The Thirty-Fifth {AAAI} Conference on Artificial Intelligence, {AAAI} 2021, Virtual Conference},
volume = {35},
number = {12},
pages = {11106--11115},
publisher = {{AAAI} Press},
year = {2021},
}
```

### Contributions

Thanks to [@kashif](https://github.com/kashif) for adding this dataset.

AI搜集汇总

数据集介绍

构建方式

ETDataset/ett数据集的构建基于北京国网富达科技发展公司的实际数据，涵盖了两个变电站的电力变压器在两年内的运行数据。数据以每小时（1H）或每15分钟（15T）的频率记录，包含17,520（或70,080）个数据点。数据集的构建旨在捕捉电力变压器油温的短期和长期周期性模式、趋势以及不规则变化，以支持时间序列预测任务。

特点

该数据集的主要特点在于其高频率和长时间跨度的数据记录，能够提供丰富的时序信息。此外，数据集包含了多种协变量，如高、中、低有用负载和无用负载，这些协变量有助于模型捕捉电力变压器油温变化的多重影响因素。数据集还提供了多变量和单变量两种设置，以适应不同预测任务的需求。

使用方法

使用ETDataset/ett数据集时，用户可以通过指定数据集的变体（如h1、h2、m1、m2）和预测长度来加载特定的数据子集。例如，使用`load_dataset('ett', 'm1', multivariate=False)`可以加载单变量15分钟频率的数据集。数据集的训练、验证和测试集分别包含12个月、4个月和4个月的数据，用户可以根据需要调整预测长度，以评估模型的性能。

背景与挑战

背景概述

电力变压器温度数据集（ETDataset/ett）由周浩一及其团队与北京国网富达科技发展有限责任公司合作创建，旨在解决电力分配中的长期预测问题。该数据集包含了两年内两个变电站的电力变压器油温数据，频率为每小时或每15分钟一次，共计17,520至70,080个数据点。核心研究问题是如何在高精度下进行长期时间序列预测，以避免电力变压器的损坏和资源浪费。该数据集的发布对时间序列预测领域具有重要影响，特别是对于电力管理和设备维护领域。

当前挑战

该数据集面临的挑战主要包括：1) 长期时间序列预测的复杂性，涉及短期和长期周期性模式、趋势以及不规则模式的处理；2) 数据集构建过程中，如何确保数据的高质量和一致性，特别是在多变量时间序列分析中，特征的选择和处理尤为关键；3) 社会影响方面，准确预测电力需求和变压器状态对于电力系统的稳定运行至关重要，任何预测误差都可能导致严重的经济和环境后果。

常用场景

经典使用场景

在电力系统管理领域，ETDataset/ett数据集以其独特的电力变压器油温时间序列数据，成为时间序列预测任务的经典范例。该数据集通过整合短期周期模式、长期周期模式、长期趋势及多种不规则模式，为研究人员提供了一个全面且复杂的时间序列分析平台。其经典使用场景包括但不限于：利用多变量时间序列预测技术，对电力变压器的油温进行精确预测，从而优化电力分配策略，减少不必要的能源浪费和设备损耗。

解决学术问题

ETDataset/ett数据集在学术研究中解决了电力系统管理中的一个关键问题：即如何基于超长期的真实世界数据进行高精度的时间序列预测。传统的预测方法往往依赖于经验数据，导致预测结果偏高，造成能源和设备的浪费。该数据集通过提供详尽的电力变压器油温数据，使得研究人员能够开发和验证更为精确的时间序列预测模型，从而在学术界推动了时间序列分析技术的发展，并为电力系统的优化管理提供了科学依据。

衍生相关工作

ETDataset/ett数据集的发布催生了一系列相关研究工作，特别是在时间序列预测和电力系统优化领域。例如，基于该数据集的研究论文《Informer: Beyond Efficient Transformer for Long Sequence Time-Series Forecasting》提出了一种新型的时间序列预测模型，显著提升了长期预测的准确性。此外，该数据集还激发了多变量时间序列分析方法的研究，推动了深度学习技术在电力系统管理中的应用。这些衍生工作不仅丰富了时间序列分析的理论体系，也为实际应用提供了新的技术手段。