This topic lists and describes the fundamental concepts underlying Energy Load Forecasting. If you are interested in learning about a concept that is not defined here or if you are interested in what a concept means in a particular context, click on the see also link or search for it in a specific ELF topic.
See the following subsections below for more information:
Aggregation from Children is one of several value assignment types that can be selected to determine how a meter or meter group receives its forecasted values.
If a meter group is not directly forecasting its own forecast value, it can be assigned a value based on an aggregation of the forecasts of its child meters. This may be useful in displaying a forecast value for a logical grouping that does not have any meaningful associated input data, and thus no calculated forecast. An aggregation that is representative of child meter or meter groups may be useful for reporting purposes.
Group aggregation is made in the form of an arithmetic expression. The operands of the expression can be a constant value or a meter or meter group facility tags. In the case of a facility tag, the net forecast value for the meter or meter group is substituted. Any meter or meter group facility tags contained in the expression must be child facilities of the target group.
Grouping and aggregation are methods that can be used to create a multi-directional meter representation. For example, a group can be created with a receipt and delivery meter as children, and an aggregation can be setup for the group that subtracts the receipt meter’s forecast from the delivery meter’s forecast.
See Configuring Aggregation From Children for more information.
Also called the “momentum parameter,” this decimal value is used during the back-propagation step of the training algorithm to help minimize the negative impact of outlying training input data. The value should be between 0.0 and 1.0 exclusive. Using an Alpha value close to 1.0 allows the neural network to make reasonably large weight adjustments as long as the corrections are in the same general direction for several patterns, while using a small learning rate to prevent a large response to the error from any one training pattern.
See Configuring Tuning Parameters for more information.
Assignment from Parent is one of several value assignment types that can be selected to determine how a meter or meter group receives its forecasted values.
A child meter or meter group can be allocated a gas load forecast value from its parent meter instead of directly forecasting its own value. In some situations it might be more useful to perform forecasting at a group level, and then assign forecast values down to child meters.
To assign a forecast value to a child meter from a parent meter group, assignment can be set using one of two assignment methods:
See Configuring Assignment From Parent for more information.
Balancing Redistribution is one of several value assignment types that can be selected to determine how a meter or meter group receives its forecasted values.
Use Balancing Redistribution to conditionally distribute a calculated forecast value for a group of meters based on whether the difference of two summed sets of meters is positive or negative. The two sets of meters represent a Sum of Receipts ( S Receipts ) and a Sum of Deliveries ( S Deliveries ). For example, this method of redistribution can assist you to make decisions about moving gas to and from storage as needed, or balancing gas receipts and deliveries across a pipeline.
The distribution supports both a configured percentage assignment value (Fixed Quantity Assignment method) and a historical time offset forecast profile method (Balance from Historical Profile Assignment method). Redistributions can be configured to occur at various points during the energy load forecasting process: after all direct forecasting, after all assignments from parent groups, or after all group aggregation.
See Configuring Balancing Redistribution for more information.
Categorical input item types must be accompanied by a table of allowable values that are valid values for a data item. Categorical data values have a list of allowable values specified for each defined input data item type. The allowable values must be 16 characters or less and will be stored in “String Input” Point Data Type points.
See Configuring Input Data Items for more information.
Direct Forecast is one of several value assignment types that can be selected to determine how a meter or meter group receives its forecasted values.
Direct forecast is the process by which a meter or meter group’s gross forecast value is generated by a neural network. Once the input data has been retrieved and manipulated as described above, the normalized inputs are fed into the associated neural network engine. The resulting output of the calculation will be the gross forecast value.
See Configuring Meter/Meter Group Assignment Settings for more information.
The energy content factor value used to convert a meter or meter group’s defined units to and from valid Energy and Volume units for the historical and forecast gas load input data. This value can be specified in the import file for meter and meter group entities, or configured manually. If the units are not provided, recognized or cannot be converted, the imported values will be rejected. If not included for an hour, the meter or meter group will use the default energy content.
See Configuring System Settings and Energy Content for more information.
The training iterations allowed in the training session for the neural network. One epoch is all hours for all inputs. If the Target Tolerance was not achieved for a particular training cycle, consider raising the target tolerance or increasing the number of maximum epochs.
See Configuring Tuning Parameters and View Training Results for more information.
Expression Redistribution is one of several value assignment types that can be selected to determine how a meter or meter group receives its forecasted values.
Expression redistribution allows a single meter or meter group to retrieve its gross energy load forecast value from another meter or meter group’s net forecast value. For example, this can be useful when a newly created meter is not ready to have an associated trained neural network. For the time being, it can receive a forecast value from a meter in a similar weather location. Expression redistribution configuration takes the form of an arithmetic expression.
Redistributions can be configured to occur at various points during the energy load forecasting process: after all direct forecasting, after all assignments from parent groups, or after all group aggregation.
See Configuring Expression Redistribution for more information.
The ID of the unique external ID value provided in the import file that links to a meter or meter group facility record.
See Configuring Meter/Meter Group General Settings and Meter for more information.
Before training occurs, the data that is designated as inputs to the training process must be filled if necessary. As each hour is imported into the system, filling rules are applied first, followed by validation rules.
Data filling settings must be associated with a particular Input Item Type (Ordinal or Categorical). If provided, the Input data filling parameters will be used to fill in the entry for an hourly interval when the hourly interval is missing from the import file. Missing hours will be filled with a placeholder with no value by the import process.
The import process detects missing hours by using the defined beginning and ending range in the import file. The adding of placeholder values during the import ensures a continuous stream of hourly data for the import data range. However, it is possible that training time ranges may specify a broader range for which there may be missing input data. Data filling will also occur for any additional missing hours that are specified in the training input ranges.
Values that are filled will be modified in the VHS. To indicate that the value was edited due to data filling, the ELF driver will flag a user bit on the stored entry for the hour.
The data filling methods for Ordinal input item types are No Action, Default Value, Previous Value, and Linear Fill. The data validation methods for Categorical input item types are No Action, Default Value, and Previous Value.
See Configuring Data Filling Rules for more information.
It is possible to override forecasted hourly values with externally provided forecast values, so that the best gas load values for meters or meter groups with a known gas load schedule are used as the forecast. Two override sources are supported (primary and secondary), one each for Energy and Volume values, to allow a staged override of forecast values.
A forecast value can be prioritized and configured to indicate the amount of time to override with the lower priority source (secondary), after the higher priority source (primary) has been used for a specified amount of time.
See Configuring Value Adjustment for more information.
The Energy Load Forecasting process utilizes a trained neural network definition to generate energy load forecast values for its associated meter or meter group. A forecasting task is completed as an entire body of work for all effective neural network definitions.
See Energy Load Forecasting for more information.
In order to forecast you must import historical and/or forecast gas load data into the system, which the ELF system uses to generate energy load forecast values for an associated meter or meter group.
See Understanding the CygNet ELF Import Schema for more information about preparing the gas load data import files.
For all gas load values imported or generated by the system, the gas load type indicates if the values will be stored in points configured in Energy units, Volume units or Both in two separate points; one with Energy units, one with Volume units for meter and meter group facilities.
See Configuring System Settings for more information.
See Rollup Forecast.
The mechanism for importing external input data for forecasting is via an XML file, which must be placed in an import folder accessible to the ELF system. There are four ELF file import types: historical gas load data, forecast gas load data, historical input data, and forecasted input data and there are there three types of folders:
See Configuring File Import Settings for more information.
In order to forecast you must import historical and/or forecast input data into the system, which the ELF system uses to generate energy load forecast values for an associated meter or meter group.
See Understanding the CygNet ELF Import Schema for more information about preparing the input data import files.
External historical input data and forecasted input data must be imported into the ELF system in order to train a neural network and generate energy load forecast values for an associated meter or meter group. These are known as Input Items.
Examples of Input Items are weather conditions (temperature, humidity, wind speed, heating degree day, cooling degree day, etc)and calendar items (time of day, day of week/month/year). An example of a non-weather Input Item would be an economic factor.
See Configuring Input Items for more information.
Input items are further organized by category. An example of an Input Item Category might be Weather.
See Configuring Input Items for more information.
Input items categories are further organized by type. There are two types of input items: Ordinal and Categorical
See Configuring Input Items for more information.
The terms “Input Owner” and “Record Owner” refer to an entity or facility that “owns” a set of input data for a given period. For example, several hours of weather input data may belong to (be owned by) the same regional weather station. The input record owner entities are specified in the import file or entered manually, and are represented as facilities in the configured CygNet FAC service. Facility type attributes are used to store associated parent and child relationships between Record Owners and Record Owner Groups.
The terms “Input Owner”and “Record Owner” are used interchangeably throughout the ELF system and in this documentation.
See Configuring Input Items, Record Owners, and Record Owner for more information.
A "lag effect" is a time offset, into the past, from which to read input data values. For example, an input data item of "ambient temperature" may be specified as input to a neural network definition causing its historical values to be used for neural network training and its forecast values to be used for forecasting. However, the ability to add an additional input data item, also associated with "ambient temperature" values but given a lag effect offset (7 days, for example), must also be provided. Providing a time lagged input value can provide significant improvements to the forecasting accuracy in the case when an input data item’s values have some correlation to what the value was in the recent past.
See Configuring Input Data Items for more information.
A decimal value (between 0.0 and 1.0 exclusive) used during the back-propagation step of the training algorithm to determine what fraction of the calculated error from each training pattern is applied as a correction to the weights of neural network being trained. A value close to 1.0 has the potential of converging to the target tolerance more quickly but also may result in an over-correction that causes the training process to fail to converge. Values between 0.1 and 0.2 are typical for yielding a successful training with a reasonable number of epochs.
See Configuring Tuning Parameters for more information.
The term "Meter" and "Meter Group" refer to the basic entity or node within the ELF system for which energy load forecast values are generated. The meter entities are specified in the import file or entered manually, and are represented as facilities in the configured CygNet FAC service. Facility type attributes are used to store associated parent and child relationships between Meters and Meter Groups.
See Configuring Meters and Input Owners, Meters, Meter for more information.
The ELF system allows for the configuration of up to 10 meter hierarchies, which are logical groupings of facilities for the purpose of forecast value assignment. Examples of different meter hierarchies may include one for Navigation, one for Sales, one for Geographic Location. A single facility (meter or meter group) can have at most 10 parents and therefore be part of up to 10 distinct meter hierarchies.
For example, multiple hierarchies may exist to allow for following circumstance: One meter hierarchy requires that direct forecast values be rolled down from a parent and split up amongst its children. Then those forecast values of the children are aggregated and rolled up to a parent in a different meter hierarchy.
Any facility (meter or meter group) in the ELF system may only get its forecast values from one source.
Hierarchy membership, parents, and children, and their related forecast assignment rules are configured on the Meter/Groups Settings Assignments dialog box.
A forecasting technique used to model the very complex relationships between known causes such as weather and calendar and their affect on customer demand for natural gas. This forecasting technique is called “feed-forward, back-propagation neural networks” or “neural networks” for short. The Estimated Load Forecasting system uses neural networks to generate the estimated customer demand (gas load) for a natural gas pipeline.
See neural network training below.
The neural network definition contains the primary set of parameters required for configuring a neural network, including those that identify all input data required to both successfully train the neural network and forecast gas load estimates for a single meter or meter group
See Configuring Neural Network Definition Settings for more information.
A pre-formatted neural network example on which to base a neural network definition. The neural network template encapsulates all of the non-meter-specific parameters required to allow the neural network to be trained and used for forecasting. Neural network templates are available to minimize duplication of configuration and effort.
See Configuring Neural Network Template Settings for more information.
The process of training a neural network requires the configuration of the neural network to produce sufficiently accurate estimates into the future based upon forecasted input data. The training algorithm is such that the neural network "learns" the relationships between the known historical inputs and their affect on the known historical outputs allowing it to predict future outputs based upon reasonable forecasted values.
Se Neural Network Training for more information.
Ordinal input item types can be assigned any analog numerical value. Ordinal data values must be 16 characters or less and will be stored in “Analog Input” Point Data Type points.
See Configuring Input Data Items for more information.
The terms “Record Owner” and “Input Owner” refer to an entity or facility that “owns” a set of input data for a given period. For example, several hours of weather input data may belong to (be owned by) the same regional weather station. The input record owner entities are specified in the import file or entered manually, and are represented as facilities in the configured CygNet FAC service. Facility type attributes are used to store associated parent and child relationships between Record Owners and Record Owner Groups.
The terms “Record Owner” and “Input Owner” are used interchangeably throughout the ELF system and in this documentation.
See Configuring Input Items, Record Owners, and Record Owner for more information.
Redistributions allow a meter or meter group to retrieve its gross forecast value from another meter or meter group’s net forecast value. For example, this can be useful when a newly created meter is not ready to have an associated trained neural network. For the time being, it can receive a forecast value from a meter in a similar weather location.
The ELF system support two types of redistribution: by expression and balancing.
See Energy Load Forecasting for more information.
Rollup Forecast is one of several value assignment types that can be selected to determine how a meter or meter group receives its forecasted values.
Rollup forecast is the process where a meter group's gross forecast values are assigned based on the rollup of the actual historical gas load values of the group's children.
See Viewing Historical Value Rollup Rules for more information.
The ID of the target meter or meter group from which historical and forecast input data will be retrieved for training and/or forecasting purposes.
See Configuring Neural Network Definition Settings for more information.
The accuracy threshold required to determine that the training process for the neural network was successful.
See Configuring Tuning Parameters and View Training Results for more information.
The accuracy threshold achieved at the completion of the training process for the neural network. If this number is less than the Target Tolerance, then the training process is considered to have met its training target goal.
See View Training Results for more information.
Before training occurs, the data that is designated as inputs to the training process must be validated. As each hour is imported into the system, filling rules are applied first, followed by validation rules.
Data validation settings must be associated with a particular Input Item Type (Ordinal or Categorical). If provided, the validation settings are used to check each input value. If a value does not comply with the validation settings, it will not be used for the corresponding hour interval. Instead, a value will be assigned for the hour based on the configured data filling settings.
The data validation methods for Ordinal data item types are Out-of-Range Validation and Deviation Validation. The data validation methods for Categorical data item types are Data Generation, and Default Value.
See Configuring Data Validation Rules for more information.
A value adjustment will occur whenever a meter or meter group’s gross point value is set by any of the defined Value Assignment Type methods. The adjustment is applied to the gross forecast value immediately upon the gross value being assigned, and output to the meter or meter group’s net forecast point. This will ensure that all adjustments are completed before redistributions are processed. Three adjustment types are available:
If no adjustments are currently active, the default action will be to copy the gross forecast to the net forecast value. The default adjustment rule will never expire, unless the end time is bounded.
See Configuring Value Adjustment for more information.
Value assignment indicate the method by which a meter or meter group can receive its forecasted values. Any node (meter or meter group) may only get its values from one source. The method of assignment is indicated by the Value Assignment Type attribute for the meter or meter group. The Value Assignment types in the ELF system areas follows:
See Configuring Meter/Meter Group Assignment Settings for more information.
Variance analysis compares the actual historical values to the forecasted values so that neural network definitions that are not performing well can be identified and retrained with the assumption that future estimates will improve. The analysis process runs at a time that is after the forecasted time interval elapses and the actual output values become known. This allows the variance between estimated and actual output values to be determined. If the variance is too high, retrain the neural network with an improved set of historical input and output values.
Variance analysis processing can add retraining tasks to the training priority queue for neural network definitions whose forecasts are not within an acceptable range of the actual energy load. The process can be triggered on demand for all qualified neural network definitions by a direct user command or scheduled through the MSS.
See Forecast Variance Analysis for more information.
The minimum calculated variance threshold that triggers the retraining of a neural network definition. For example, a value of 0.7 means that the forecast is 30% accurate. This value may be overridden at the meter or meter group facility.
See Configuring System Settings and Configuring Tuning Parameters for more information.
Prior to training a neural network, the internal weights of the neural network must be initialized to random values. The weight assignment values allow some control over how the internal weights are assigned so that optimal settings may be determined and saved. The weight assignment can be either an arbitrary seed or a seed value.
See Configuring Tuning Parameters for more information.
|
|
|
