How our forecasting models work

Building models

When building load forecasting models for a given territory, we start with the historical load data of that territory. That will need to be provided to our team for analysis and some basic back-and-forth as we understand the data.

We will also run a process called a Weather Station Optimization, or WSO, to determine the best set of nearby weather station data to use when forecasting demand in a given territory.

To read more about historical data required to build forecasting models for a territory, see Historical Data from the "All About Data" page.

Ongoing inputs to the models

Load values

Depending on the specific model type that you are licensing, MCast™ utilizes different load values as inputs to your model

2-Day-Ago Flow Models use the most recently entered observed loads from 9-2 days ago. This is the most common model type for customers who have data available from 2 gas days ago
1-Day-Ago Flow Models use the most recently entered observed loads from 9-1 day ago. This is a special model type enabled for customers that receive updates from the previous gas day during today.
No Flow Models do not use any observed loads as a direct input into the model. However, it is important to understand that MCast's tuning process considers the results of any entered values to calibrate the weighting between different models

Remember that, even though observed loads may not be directly fed into a "No Flow" model, their impact is still significant during the calibration process.

To read more about submitting your observed load data to run MCast™, see Ongoing Load Data from the "All About Data" page.

Weather values

All MCast™ models require ongoing weather data actuals and weather data forecasts to be provided in an hourly format in order to generate new load forecasts. Because energy demand is highly sensitive to weather, changes in the weather forecasts are the most common reasons that customers experience changes to their load forecasts.

To read more about the implications to, and options for, sourcing ongoing weather data for MCast™, see Ongoing Weather Data from the "All About Data" page.

Intraday Models

An intraday model is a specialized model that only runs during specific periods of the day. These models have distinct data requirements based on the data available during designated time frames.

Example of running 2 intraday models

2-Day-Ago Flow Model (6:00 AM - 10: 59 AM)

During the morning hours, MCast™ runs a 2-day-ago flow model. The resulting forecasts can be tracked and analyzed at any time on the Performance Page using the Intraday Forecast selector. Additionally, you have the flexibility to run multiple forecasts during this time period. The forecast pinning feature can be used to help to distinguish between forecasts run during this period.

1-Day-Ago Flow Model (11:00 AM - 5:59 AM)

At 11 AM, MCast™ switches to using a 1-day-ago flow model. Again the forecasts generated during this time can be evaluated through the Performance Page, independent of the results of the 2-Day-Ago flow model. Notably though, new forecasts can not be generated during this period if you have not uploaded the observed load from the prior gasday.

Note

If you have any questions or want to make any changes to your company's usage of intraday forecasts please reach out to support@marquetteenergyanalytics.com

Accounting for holidays in daily demand

Energy demand often differs on holidays, as people's schedules change. Our forecasting models adjust for this change in pattern using two methods: Day-of-the-week adjustment and Flow adjustment.

Day-of-the-week Adjustment: A part of our forecasting models considers which day of the week a given day is to adjust its forecast. On a holiday, the daily load often looks more like a weekend day than a weekday. Because of this, the first adjustment our forecasting models make on holidays is to treat them as though they were a different day of the week.

Example

Thanksgiving Day acts more like a Saturday than a Thursday, so we set it to Saturday. In addition, the day before Thanksgiving Day acts more like a Friday than a Wednesday, so we set it to a Friday.

Many holidays fall on Mondays. For these days, we set the day before the holiday to act like a Saturday instead of a Sunday, and set the holiday itself to act like a Sunday instead of a Monday. This is because Sunday will not have the typical Monday morning startup so it acts more like a Saturday and the Monday will have the beginning of the week startup which makes it behave more like a Sunday.

Flow Adjustment: Our forecasting models make a secondary adjustment for the holidays. The linear regression component models are built with day-of-the-week adjustments for the holidays. When these models are trained, we calculate the mean residual errors for each holiday. The forecasting models, in addition to the day-of-the-week adjustment, use this residual error when forecasting a holiday by subtracting it from the component model forecasts.

This flow adjustment is used on minor and non-holidays too, when no day-of-the-week adjustment is made.

Example

On Presidents Day and Veterans Day, which are federal and bank holidays, but most commercial and industrial customers are unaffected, a day-of-the-week adjustment is too great. On these days we make only the flow adjustment. This flow adjustment is also done on DST-to-ST and ST-to-DST days.

Post-holiday Adjustment: Important inputs into the forecasting models are the actual demands on recent days. When a holiday is in these recent days, the low demand on the holiday will bias the forecast low. To overcome this bias, we adjust the actual holiday days' demand to estimate what the demand would have been had these days not been holidays. This adjusted demand is used by the forecasting models.

Accounting for Daylight Savings Time

Modeling Daylight Savings Time

Flow Adjustment: MCast™ treats the transitions from Standard Time to Daylight Savings Time and back as special cases of "holidays" and makes a flow adjustment. When your linear regression component models are built, our team calculates the mean residual errors for each "holiday." Our forecasting models use this residual error when forecasting on these holidays by subtracting it from the component model forecasts.

Reporting your data for Daylight Savings Time

We support 2 conventional ways of handling Daylight Savings Time adjustments for hourly weather values and hourly load data: data produced in Clock Time or in Standard Time,

Data in Clock Time follows the local timezone's clock, accounting for the Daylight Savings adjustments with a 23-hour day in the spring and a 25-hour day in the fall. On the 23-hour day, the 2:00 hour is removed from the file. On the 25-hour day, the 2:00 hour appears twice in the file.
Data in Standard Time follows the local timezone as if Daylight Savings time is never in use. This maintains a 24-hour day for every day of the year, so when DST is in-use (March-November) the data being delivered is just an hour shifted from clock time.

Please work with a member of the Marquette Energy Analytics support team to ensure that we know if your data is delivered in Clock Time or Standard Time.

Warning

If your data is not correctly formatted as either Clock Time or Standard Time, the accuracy of your hourly forecasts will be negatively affected.

Modeling the First Cold Days of the Heating Season (webinar)

In the video below, our team discusses a common question that we receive - How do MCast™ forecasting models handle the first cold days of the heating season? You'll hear our team walk through a sample model one parameter at a time, and point out how the different parameters and data treatments factor into our forecasting on those first cold days in MCast™.

Timestamps

0:00 - Graphing demand vs. flow data
0:30 - Defining the "First Cold Days"
1:26 - Fitting a 2-parameter model
2:23 - Adding a 3rd parameter: HDD55
3:55 - Adding a 4th parameter: CDD
5:15 - Adding a 5th parameter: yesterday's HDD
6:22 - Understanding the 5-parameter model
8:15 - Other data treatments

Common Questions

How do the daily models track changing load?

There are two main ways the MCast™ daily models track and respond to changes in demand across an area.

All MCast™ daily models, except for the No Flow models, are autoregressive. This means that recent loads are inputs into the model. More information on that can be found here.
All daily models use a tuning mechanism that uses recent errors to adjust the model. This is designed to track changing customer bases and usage patterns to respond to changes over time.

It is important to note that both of these mechanisms rely upon the observed values being entered into MCast™. If erroneous/bad data is uploaded to MCast™, the models may see this as a pattern or behavior change and start adjusting their forecasts accordingly. This is why it is important to always update MCast™ with the most recent, reliable observed values.

Additionally, MCast™ daily models are retrained once every year, and during this process, our data science team analyzes the data for changing patterns and outliers and delivers a full new set of models

How do the hourly models track changing load?

There are two main ways the MCast™ hourly models track and respond to changes in demand across an area.

All MCast™ hourly models are autoregressive. This means that recent loads are inputs into the model.
MCast™ hourly models are retrained every three months. During this process, our data science team analyzes the data for any outliers or changing patterns and delivers a full new set of models.

What are the units of my MCast™ demand forecast?

MCast™ can forecast in any unit of load, flow, or demand because the forecasts it produces are effectively unitless. The units of data that are input to MCast™ will be the same as the units of data that are output. For this reason, it is crucial that you are sending consistent units to the Marquette Energy Analytics team.

Please ensure that you discuss with your Marquette Energy Analytics contact what units you will be providing data in.

Unit Examples

If you provided your historical load data in decatherms, then the models that our team builds will be optimized to forecast on decatherms. Those models will expect to receive data in decatherms over time, and similarly will produce their estimates in decatherms.

If you provided your historical load data in MCF, then the models that our team builds will be optimized to forecast on MCF. Those models will expect to receive data in MCF over time, and similarly will produce their estimates in MCF.