{ "id": 118753, "title": "Problem with the dataset - Measurement Units", "url": "/competitions/ashrae-energy-prediction/discussion/118753", "author_name": "", "post_date": "2019-11-24T07:24:57.181536600Z", "votes": 22, "comment_count": 12, "views": 0, "content": "

Now that there have been various site's leaked. We now know the units (per meter) each (publically known) site is reporting at. \n| Site | Location | Electric | Chilled | Steam | Hot\n| --- | --- | --- | --- | --- | --- |\n| 0 | UCF | kBTU | kBTU | n/a | n/a |\n| 1 | UCL | kW | - | - | - |\n| 2 | ASU | kW | Tons | n/a | mmBTU |\n| 4 | UCB | kW | n/a | lbs | n/a |\n| 15 | Cornell | kW | Tons | lbs | n/a |

\n\n

Okay, so 3 of the known sites (2,4,15) appear to be reporting in native units, while UCF is reporting is kBTU.

\n\n

The big issue is the fact that we don't know about the units each site's meter's are recorded at. This in my opinion is a problem. It's possible that UCF is the only site out of 16 sites that is reporting differently, however, I would rather have peace of mind knowing that the rest of the sites share the consistency seen above rather than guessing.

\n\n

If the units were provided to us, we could create better models. If you're building 2380 models, one per building's meter, okay, no big deal. But if you're building a single model and the target variables are in different units, then it likely that you're underfitting (unknowingly). Why? because you'll be training your model in different units for the same meter type. This is not ideal.

\n\n

To bring the point home, imagine if your dataset was stock prices and some companies were reported by the data source was reporting in market capital instead of price per share. You would have different measurements for your stocks. This is what this dataset looks like.

", "messages": [ { "id": "680190", "postDate": "11/24/2019 07:24:57", "content": "

Now that there have been various site's leaked. We now know the units (per meter) each (publically known) site is reporting at. \n| Site | Location | Electric | Chilled | Steam | Hot\n| --- | --- | --- | --- | --- | --- |\n| 0 | UCF | kBTU | kBTU | n/a | n/a |\n| 1 | UCL | kW | - | - | - |\n| 2 | ASU | kW | Tons | n/a | mmBTU |\n| 4 | UCB | kW | n/a | lbs | n/a |\n| 15 | Cornell | kW | Tons | lbs | n/a |

\n\n

Okay, so 3 of the known sites (2,4,15) appear to be reporting in native units, while UCF is reporting is kBTU.

\n\n

The big issue is the fact that we don't know about the units each site's meter's are recorded at. This in my opinion is a problem. It's possible that UCF is the only site out of 16 sites that is reporting differently, however, I would rather have peace of mind knowing that the rest of the sites share the consistency seen above rather than guessing.

\n\n

If the units were provided to us, we could create better models. If you're building 2380 models, one per building's meter, okay, no big deal. But if you're building a single model and the target variables are in different units, then it likely that you're underfitting (unknowingly). Why? because you'll be training your model in different units for the same meter type. This is not ideal.

\n\n

To bring the point home, imagine if your dataset was stock prices and some companies were reported by the data source was reporting in market capital instead of price per share. You would have different measurements for your stocks. This is what this dataset looks like.

", "rawMarkdown": "Now that there have been various site's leaked. We now know the units (per meter) each (publically known) site is reporting at. \n| Site | Location | Electric | Chilled | Steam | Hot\n| --- | --- | --- | --- | --- | --- |\n| 0 | [UCF](https://www.oeis.ucf.edu/buildings) | kBTU | kBTU | n/a | n/a |\n| 1 | [UCL](https://platform.carbonculture.net/communities/ucl/30/apps/assets/list/place/) | kW | - | - | - |\n| 2 | [ASU](http://cm.asu.edu/) | kW | Tons | n/a | mmBTU |\n| 4 | [UCB](https://engagementdashboard.com/ucb/ucb/) | kW | n/a | lbs | n/a |\n| 15 | [Cornell](http://portal.emcs.cornell.edu/) | kW | Tons | lbs | n/a |\n\nOkay, so 3 of the known sites (2,4,15) appear to be reporting in native units, while UCF is reporting is kBTU.\n\nThe big issue is the fact that we don't know about the units each site's meter's are recorded at. This in my opinion is a problem. It's possible that UCF is the only site out of 16 sites that is reporting differently, however, I would rather have peace of mind knowing that the rest of the sites share the consistency seen above rather than guessing.\n\nIf the units were provided to us, we could create better models. If you're building 2380 models, one per building's meter, okay, no big deal. But if you're building a single model and the target variables are in different units, then it likely that you're underfitting (unknowingly). Why? because you'll be training your model in different units for the same meter type. This is not ideal.\n\nTo bring the point home, imagine if your dataset was stock prices and some companies were reported by the data source was reporting in market capital instead of price per share. You would have different measurements for your stocks. This is what this dataset looks like.", "votes": null }, { "id": "680336", "postDate": "11/24/2019 13:37:43", "content": "

More reasons why this entire competition is a joke. As an ASHRAE Life Member, I'm embarrassed.

", "rawMarkdown": "More reasons why this entire competition is a joke. As an ASHRAE Life Member, I'm embarrassed.", "votes": null }, { "id": "681149", "postDate": "11/25/2019 18:30:08", "content": "

Per the data description: meter_reading - The target variable. Energy consumption in kWh (or equivalent).

\n\n

We standardized the units before posting the data:\n- Asking 2,000 people to repeat the same preprocessing steps did not seem like it would be a fun addition to the competition\n- Not everyone reads the data description

", "rawMarkdown": "Per the data description: `meter_reading - The target variable. Energy consumption in kWh (or equivalent).`\n\nWe standardized the units before posting the data:\n- Asking 2,000 people to repeat the same preprocessing steps did not seem like it would be a fun addition to the competition\n- Not everyone reads the data description", "votes": null }, { "id": "681199", "postDate": "11/25/2019 20:25:18", "content": "

Understood, however UCF's website is reporting in kBTU for electricity. Can't really argue with that since the values in 2016 from UCF's website and the training dataset match up. Unless of course the website is wrong. I'm just going to take the L here and accept that the website is wrong.

", "rawMarkdown": "Understood, however UCF's website is reporting in kBTU for electricity. Can't really argue with that since the values in 2016 from UCF's website and the training dataset match up. Unless of course the website is wrong. I'm just going to take the L here and accept that the website is wrong.", "votes": null }, { "id": "681261", "postDate": "11/25/2019 23:07:58", "content": "

Where is the data description? I looked in the data section and column information, and it doesn't indicate units.

\n\n

EDIT: Ah, drop down arrow. Didn't notice it before. I retract my comment. Thanks for the response.

", "rawMarkdown": "Where is the data description? I looked in the data section and column information, and it doesn't indicate units.\n\nEDIT: Ah, drop down arrow. Didn't notice it before. I retract my comment. Thanks for the response.", "votes": null }, { "id": "682638", "postDate": "11/27/2019 17:48:39", "content": "

@teeyee314 I looked into this further and confirmed that the electric meter readings for site 0 were not properly converted from kBTU to kWh. I've made a general post here discussing the issue. Thank you for the catch!

", "rawMarkdown": "teeyee314 I looked into this further and confirmed that the electric meter readings for site 0 were not properly converted from kBTU to kWh. I've made [a general post here](https://www.kaggle.com/c/ashrae-energy-prediction/discussion/119261) discussing the issue. Thank you for the catch!", "votes": null }, { "id": "682759", "postDate": "11/27/2019 20:09:56", "content": "

Another inconsistency is site 1 readings are aggregated on hourly sum and site 15 readings are aggregated on hourly mean

", "rawMarkdown": "Another inconsistency is site 1 readings are aggregated on hourly sum and site 15 readings are aggregated on hourly mean", "votes": null }, { "id": "682772", "postDate": "11/27/2019 20:28:52", "content": "

The more you know.🤔

", "rawMarkdown": "The more you know.🤔", "votes": null }, { "id": "683273", "postDate": "11/28/2019 09:04:53", "content": "

Is it just the electricity meter readings that need to be converted or do the chilled ones also need to be done? In the table above, site 0 also has different units for these.

", "rawMarkdown": "Is it just the electricity meter readings that need to be converted or do the chilled ones also need to be done? In the table above, site 0 also has different units for these.", "votes": null }, { "id": "686061", "postDate": "12/02/2019 17:43:22", "content": "

Unless I'm mistaken those are just two different ways of describing the same concept of kWh. If you run a 2 kW device for 0.5 hours, you will get 1 kWh as the result of both calculations.

", "rawMarkdown": "Unless I'm mistaken those are just two different ways of describing the same concept of kWh. If you run a 2 kW device for 0.5 hours, you will get 1 kWh as the result of both calculations.", "votes": null }, { "id": "686085", "postDate": "12/02/2019 18:48:46", "content": "

That would be correct if only we had 1 hour of meter readings

", "rawMarkdown": "That would be correct if only we had 1 hour of meter readings", "votes": null }, { "id": "686136", "postDate": "12/02/2019 20:24:12", "content": "

Depends on the frequency of the meter measurement. For example, at my institution, we push meter readings every 5 minutes. To get the energy consumption, the calculation would be (KW(t)-KW(t-1)) x interval(minutes)/60 (minutes/hour). The KWh would be the sum between any two intervals. If the frequency is once per hour (which the competition standardized on), then KWh = (KW(t)-KW(t-1)) x 1 hour. If, however, you're looking for demand (KW), then for a measurement frequency of one hour, it would be the mean of readings between t and t-1. The measurements aren't actually in KWh, they're in KW.

", "rawMarkdown": "Depends on the frequency of the meter measurement. For example, at my institution, we push meter readings every 5 minutes. To get the energy consumption, the calculation would be (KW(t)-KW(t-1)) x interval(minutes)/60 (minutes/hour). The KWh would be the sum between any two intervals. If the frequency is once per hour (which the competition standardized on), then KWh = (KW(t)-KW(t-1)) x 1 hour. If, however, you're looking for demand (KW), then for a measurement frequency of one hour, it would be the mean of readings between t and t-1. The measurements aren't actually in KWh, they're in KW.", "votes": null }, { "id": "690691", "postDate": "12/09/2019 00:29:19", "content": "

The data is highly valuable, but the inconsistency is literally frustrating. Now, most of the time wasted here is just about web scraping instead of working out a useful model.

", "rawMarkdown": "The data is highly valuable, but the inconsistency is literally frustrating. Now, most of the time wasted here is just about web scraping instead of working out a useful model.", "votes": null } ], "comments": [ { "id": 680336, "author_name": "ljschuster", "author_url": "", "post_date": "11/24/2019 13:37:43", "content": "

More reasons why this entire competition is a joke. As an ASHRAE Life Member, I'm embarrassed.

", "votes": null, "replies": [ { "id": 690691, "author_name": "smerllo", "author_url": "", "post_date": "12/09/2019 00:29:19", "content": "

The data is highly valuable, but the inconsistency is literally frustrating. Now, most of the time wasted here is just about web scraping instead of working out a useful model.

", "votes": null, "replies": [] } ] }, { "id": 681149, "author_name": "sohier", "author_url": "", "post_date": "11/25/2019 18:30:08", "content": "

Per the data description: meter_reading - The target variable. Energy consumption in kWh (or equivalent).

\n\n

We standardized the units before posting the data:\n- Asking 2,000 people to repeat the same preprocessing steps did not seem like it would be a fun addition to the competition\n- Not everyone reads the data description

", "votes": null, "replies": [ { "id": 681199, "author_name": "teeyee314", "author_url": "", "post_date": "11/25/2019 20:25:18", "content": "

Understood, however UCF's website is reporting in kBTU for electricity. Can't really argue with that since the values in 2016 from UCF's website and the training dataset match up. Unless of course the website is wrong. I'm just going to take the L here and accept that the website is wrong.

", "votes": null, "replies": [] }, { "id": 681261, "author_name": "ljschuster", "author_url": "", "post_date": "11/25/2019 23:07:58", "content": "

Where is the data description? I looked in the data section and column information, and it doesn't indicate units.

\n\n

EDIT: Ah, drop down arrow. Didn't notice it before. I retract my comment. Thanks for the response.

", "votes": null, "replies": [] }, { "id": 682638, "author_name": "sohier", "author_url": "", "post_date": "11/27/2019 17:48:39", "content": "

@teeyee314 I looked into this further and confirmed that the electric meter readings for site 0 were not properly converted from kBTU to kWh. I've made a general post here discussing the issue. Thank you for the catch!

", "votes": null, "replies": [] }, { "id": 683273, "author_name": "steveroberts", "author_url": "", "post_date": "11/28/2019 09:04:53", "content": "

Is it just the electricity meter readings that need to be converted or do the chilled ones also need to be done? In the table above, site 0 also has different units for these.

", "votes": null, "replies": [] } ] }, { "id": 682759, "author_name": "gunesevitan", "author_url": "", "post_date": "11/27/2019 20:09:56", "content": "

Another inconsistency is site 1 readings are aggregated on hourly sum and site 15 readings are aggregated on hourly mean

", "votes": null, "replies": [ { "id": 682772, "author_name": "teeyee314", "author_url": "", "post_date": "11/27/2019 20:28:52", "content": "

The more you know.🤔

", "votes": null, "replies": [] }, { "id": 686061, "author_name": "sohier", "author_url": "", "post_date": "12/02/2019 17:43:22", "content": "

Unless I'm mistaken those are just two different ways of describing the same concept of kWh. If you run a 2 kW device for 0.5 hours, you will get 1 kWh as the result of both calculations.

", "votes": null, "replies": [] }, { "id": 686085, "author_name": "gunesevitan", "author_url": "", "post_date": "12/02/2019 18:48:46", "content": "

That would be correct if only we had 1 hour of meter readings

", "votes": null, "replies": [] }, { "id": 686136, "author_name": "ljschuster", "author_url": "", "post_date": "12/02/2019 20:24:12", "content": "

Depends on the frequency of the meter measurement. For example, at my institution, we push meter readings every 5 minutes. To get the energy consumption, the calculation would be (KW(t)-KW(t-1)) x interval(minutes)/60 (minutes/hour). The KWh would be the sum between any two intervals. If the frequency is once per hour (which the competition standardized on), then KWh = (KW(t)-KW(t-1)) x 1 hour. If, however, you're looking for demand (KW), then for a measurement frequency of one hour, it would be the mean of readings between t and t-1. The measurements aren't actually in KWh, they're in KW.

", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "680190": "Now that there have been various site's leaked. We now know the units (per meter) each (publically known) site is reporting at. \n| Site | Location | Electric | Chilled | Steam | Hot\n| --- | --- | --- | --- | --- | --- |\n| 0 | [UCF](https://www.oeis.ucf.edu/buildings) | kBTU | kBTU | n/a | n/a |\n| 1 | [UCL](https://platform.carbonculture.net/communities/ucl/30/apps/assets/list/place/) | kW | - | - | - |\n| 2 | [ASU](http://cm.asu.edu/) | kW | Tons | n/a | mmBTU |\n| 4 | [UCB](https://engagementdashboard.com/ucb/ucb/) | kW | n/a | lbs | n/a |\n| 15 | [Cornell](http://portal.emcs.cornell.edu/) | kW | Tons | lbs | n/a |\n\nOkay, so 3 of the known sites (2,4,15) appear to be reporting in native units, while UCF is reporting is kBTU.\n\nThe big issue is the fact that we don't know about the units each site's meter's are recorded at. This in my opinion is a problem. It's possible that UCF is the only site out of 16 sites that is reporting differently, however, I would rather have peace of mind knowing that the rest of the sites share the consistency seen above rather than guessing.\n\nIf the units were provided to us, we could create better models. If you're building 2380 models, one per building's meter, okay, no big deal. But if you're building a single model and the target variables are in different units, then it likely that you're underfitting (unknowingly). Why? because you'll be training your model in different units for the same meter type. This is not ideal.\n\nTo bring the point home, imagine if your dataset was stock prices and some companies were reported by the data source was reporting in market capital instead of price per share. You would have different measurements for your stocks. This is what this dataset looks like.", "680336": "More reasons why this entire competition is a joke. As an ASHRAE Life Member, I'm embarrassed.", "681149": "Per the data description: `meter_reading - The target variable. Energy consumption in kWh (or equivalent).`\n\nWe standardized the units before posting the data:\n- Asking 2,000 people to repeat the same preprocessing steps did not seem like it would be a fun addition to the competition\n- Not everyone reads the data description", "681199": "Understood, however UCF's website is reporting in kBTU for electricity. Can't really argue with that since the values in 2016 from UCF's website and the training dataset match up. Unless of course the website is wrong. I'm just going to take the L here and accept that the website is wrong.", "681261": "Where is the data description? I looked in the data section and column information, and it doesn't indicate units.\n\nEDIT: Ah, drop down arrow. Didn't notice it before. I retract my comment. Thanks for the response.", "682638": "teeyee314 I looked into this further and confirmed that the electric meter readings for site 0 were not properly converted from kBTU to kWh. I've made [a general post here](https://www.kaggle.com/c/ashrae-energy-prediction/discussion/119261) discussing the issue. Thank you for the catch!", "682759": "Another inconsistency is site 1 readings are aggregated on hourly sum and site 15 readings are aggregated on hourly mean", "682772": "The more you know.🤔", "683273": "Is it just the electricity meter readings that need to be converted or do the chilled ones also need to be done? In the table above, site 0 also has different units for these.", "686061": "Unless I'm mistaken those are just two different ways of describing the same concept of kWh. If you run a 2 kW device for 0.5 hours, you will get 1 kWh as the result of both calculations.", "686085": "That would be correct if only we had 1 hour of meter readings", "686136": "Depends on the frequency of the meter measurement. For example, at my institution, we push meter readings every 5 minutes. To get the energy consumption, the calculation would be (KW(t)-KW(t-1)) x interval(minutes)/60 (minutes/hour). The KWh would be the sum between any two intervals. If the frequency is once per hour (which the competition standardized on), then KWh = (KW(t)-KW(t-1)) x 1 hour. If, however, you're looking for demand (KW), then for a measurement frequency of one hour, it would be the mean of readings between t and t-1. The measurements aren't actually in KWh, they're in KW.", "690691": "The data is highly valuable, but the inconsistency is literally frustrating. Now, most of the time wasted here is just about web scraping instead of working out a useful model." }, "source": "meta" }