Class TemporalFusionTransformerOptions<T>

Namespace

AiDotNet.Models.Options

Assembly

AiDotNet.dll

Configuration options for the Temporal Fusion Transformer (TFT) model.

public class TemporalFusionTransformerOptions<T> : TimeSeriesRegressionOptions<T>

Type Parameters

T

The numeric type used for calculations (typically double or float).

Inheritance

object

ModelOptions

RegressionOptions<T>

TimeSeriesRegressionOptions<T>

TemporalFusionTransformerOptions<T>

Inherited Members

TimeSeriesRegressionOptions<T>.LagOrder

TimeSeriesRegressionOptions<T>.IncludeTrend

TimeSeriesRegressionOptions<T>.SeasonalPeriod

TimeSeriesRegressionOptions<T>.AutocorrelationCorrection

TimeSeriesRegressionOptions<T>.ModelType

TimeSeriesRegressionOptions<T>.LossFunction

RegressionOptions<T>.DecompositionMethod

RegressionOptions<T>.UseIntercept

ModelOptions.Seed

object.Equals(object)

object.Equals(object, object)

object.GetHashCode()

object.GetType()

object.MemberwiseClone()

object.ReferenceEquals(object, object)

object.ToString()

Remarks

Temporal Fusion Transformer is a state-of-the-art deep learning architecture for multi-horizon forecasting. It combines high-performance multi-horizon forecasting with interpretable insights into temporal dynamics. TFT uses self-attention mechanisms to learn temporal relationships at different scales and integrates static metadata, time-varying known inputs, and time-varying unknown inputs.

For Beginners: TFT is an advanced neural network designed specifically for forecasting that can handle multiple types of input data: - Static features (e.g., store location, product category) that don't change over time - Known future inputs (e.g., holidays, promotions) that we know ahead of time - Unknown inputs (e.g., past sales) that we can only observe historically

The model uses "attention" mechanisms to focus on the most relevant time periods and features, making it both accurate and interpretable.

Constructors

TemporalFusionTransformerOptions()

Initializes a new instance of the TemporalFusionTransformerOptions<T> class.

public TemporalFusionTransformerOptions()

TemporalFusionTransformerOptions(TemporalFusionTransformerOptions<T>)

Initializes a new instance by copying from another instance.

public TemporalFusionTransformerOptions(TemporalFusionTransformerOptions<T> other)

Parameters

other TemporalFusionTransformerOptions<T>

Properties

BatchSize

Gets or sets the batch size for training.

public int BatchSize { get; set; }

Property Value

int

The batch size, defaulting to 32.

DropoutRate

Gets or sets the dropout rate for regularization.

public double DropoutRate { get; set; }

Property Value

double

The dropout rate, defaulting to 0.1.

Remarks

For Beginners: Dropout randomly ignores some neurons during training to prevent overfitting. A value of 0.1 means 10% of neurons are ignored in each training step.

Epochs

Gets or sets the number of training epochs.

public int Epochs { get; set; }

Property Value

int

The number of epochs, defaulting to 100.

ForecastHorizon

Gets or sets the forecast horizon (number of future time steps to predict).

public int ForecastHorizon { get; set; }

Property Value

int

The forecast horizon, defaulting to 6.

HiddenSize

Gets or sets the hidden state size for the model.

public int HiddenSize { get; set; }

Property Value

int

The hidden state size, defaulting to 128.

Remarks

For Beginners: This controls the capacity of the model's internal representations. Larger values allow the model to capture more complex patterns but require more memory and computation.

LearningRate

Gets or sets the learning rate for training.

public double LearningRate { get; set; }

Property Value

double

The learning rate, defaulting to 0.001.

LookbackWindow

Gets or sets the lookback window size (number of historical time steps used as input).

public int LookbackWindow { get; set; }

Property Value

int

The lookback window size, defaulting to 24.

NumAttentionHeads

Gets or sets the number of attention heads in the multi-head attention mechanism.

public int NumAttentionHeads { get; set; }

Property Value

int

The number of attention heads, defaulting to 4.

Remarks

For Beginners: Attention heads allow the model to focus on different aspects of the time series simultaneously. More heads can capture more diverse patterns.

NumLayers

Gets or sets the number of transformer layers.

public int NumLayers { get; set; }

Property Value

int

The number of layers, defaulting to 2.

QuantileLevels

Gets or sets the quantile levels for probabilistic forecasting.

public double[] QuantileLevels { get; set; }

Property Value

double[]

Array of quantile levels, defaulting to [0.1, 0.5, 0.9].

Remarks

For Beginners: Quantile forecasting provides prediction intervals. For example, [0.1, 0.5, 0.9] gives you the 10th percentile (pessimistic), median (most likely), and 90th percentile (optimistic) predictions.

StaticCovariateSize

Gets or sets the number of static covariates (features that don't change over time).

public int StaticCovariateSize { get; set; }

Property Value

int

The number of static covariates, defaulting to 0.

TimeVaryingKnownSize

Gets or sets the number of time-varying known inputs (future values that are known).

public int TimeVaryingKnownSize { get; set; }

Property Value

int

The number of time-varying known inputs, defaulting to 0.

TimeVaryingUnknownSize

Gets or sets the number of time-varying unknown inputs (past observations only).

public int TimeVaryingUnknownSize { get; set; }

Property Value

int

The number of time-varying unknown inputs, defaulting to 1.

UseVariableSelection

Gets or sets whether to use variable selection networks.

public bool UseVariableSelection { get; set; }

Property Value

bool

True to use variable selection, defaulting to true.

Remarks

For Beginners: Variable selection automatically determines which input features are most important for making predictions, improving both accuracy and interpretability.