Datasets:

MaxKJC
/

TSDBench

# ============================================================
# Task: traffic_flow_forecasting
# ============================================================
# Task objective:
# - Use a historical window of LargeST traffic_flow to predict a future window.
# - This is a typical multisensor time series forecasting task.
#
# Core semantics:
# - sample defines one sample: a decision timestamp.
# - inputs defines what the model can access: historical traffic_flow before timestamp,
#   plus future covariates known at prediction time, such as day_of_week.
# - prediction defines what to predict: the traffic_flow sequence over the future horizon.
# - leakage defines which future values must be excluded from inputs: future traffic_flow must never be an input.
# ============================================================

task:
  id: traffic_flow_forecasting # Unique task ID; keep it consistent with the directory name for automatic discovery by scripts.
  type: time_series_forecasting # Task type: time series forecasting.
  description: >
    Predict future LargeST traffic_flow from historical traffic_flow.
    Each sample corresponds to one decision timestamp, and labels are stored
    by index reference to avoid expanding all sensor-by-time targets into labels.csv.
# Whether to inherit the general processing rules in ../general_task.yaml.
inherit_general: true

sample:
  primary_key: timestamp

  # Label storage mode.
  # index_reference: labels.csv does not expand every (ID, timestamp, horizon) label;
  #                  instead, reference columns point to the underlying time series file.
  # long_table:      labels.csv stores each target value in a separate row; this is larger but easier to inspect.
  label_mode: index_reference

inputs:
  # ------------------------------------------------------------
  # Model input 1: historical traffic flow.
  # ------------------------------------------------------------
  time_series:
    - fields: traffic_flow  # Historical time series fields available to the model.
      source: traffic_flow_ca_2019 # Refers to the underlying LargeST h5 file defined in data_refs.time_series.

      window:
        length: 64 # The history window contains 672 task time steps.
        unit: 15min # Each step is 15 minutes, so the model sees the past 7 days.

      sampling:  # Optional configuration; set this here if the raw data is not sampled every 15min.
        native_rate: 15min # Native sampling resolution of the LargeST h5 file.
        task_rate: 15min # Input resolution for this task, matching the native resolution.
        downsample_method: mean # If changed to 1h, use mean to aggregate four 15min observations.
        upsample_method: ffill # If changed to a finer resolution, use forward filling.



  tabular_static:
    source: largeST_static_metadata # Refers to the static attribute table in data_refs.tabular_static.

    fields:
      - Lat
      - Lng
      - District
      - County
      - Fwy
      - Lanes
      - Type
      - Direction


  graph:
      source: road_distance_graph

prediction:
  field: traffic_flow # Target variable to predict.
  source: traffic_flow_ca_2019 # Target values come from the same LargeST historical traffic flow h5 file.

  horizon:
    length: 16 # Predict the next 16 time steps.
    unit: 15min # Each step is 15 minutes, so the forecast covers the next 4 hours.
    starts_after: timestamp # The future window starts after sample.timestamp.

  stride: 16 # Timestamps of adjacent samples are 16 task time steps apart.

  label_storage:
    ref_column: target_series_ref # Reference column in labels.csv pointing to the underlying target sequence.
    shape_column: target_tensor_shape # Column in labels.csv recording the target tensor shape, for example [num_nodes, horizon, num_fields].

leakage:

  observed_past_only_columns:
    - traffic_flow # traffic_flow can only be a historical input; future traffic_flow is the prediction target and must not be an input.



data_refs:
  time_series:
    - traffic_flow_ca_2019:
      file: All_Datasets/LargeST_ca_full/input_multimodal/time_series/traffic_flow/ca_his_2019.h5 # LargeST California 2019 historical traffic flow file.
      value_field: traffic_flow # Semantics of the corresponding variable in the h5 file.
      native_grid_minutes: 15 # The native time grid is 15 minutes.



  tabular_static:
    largeST_static_metadata:
      file: All_Datasets/LargeST_ca_full/input_multimodal/ca_meta.csv # Static sensor attribute table.
      id_column: ID

  graph:
    road_distance_graph:
      file: All_Datasets/LargeST_ca_full/input_multimodal/node_level_task_graphs/road_distance_graph/edge_infos.csv


	# ============================================================
	# Task: traffic_flow_forecasting
	# ============================================================
	# Task objective:
	# - Use a historical window of LargeST traffic_flow to predict a future window.
	# - This is a typical multisensor time series forecasting task.
	#
	# Core semantics:
	# - sample defines one sample: a decision timestamp.
	# - inputs defines what the model can access: historical traffic_flow before timestamp,
	# plus future covariates known at prediction time, such as day_of_week.
	# - prediction defines what to predict: the traffic_flow sequence over the future horizon.
	# - leakage defines which future values must be excluded from inputs: future traffic_flow must never be an input.
	# ============================================================

	task:
	id: traffic_flow_forecasting # Unique task ID; keep it consistent with the directory name for automatic discovery by scripts.
	type: time_series_forecasting # Task type: time series forecasting.
	description: >
	Predict future LargeST traffic_flow from historical traffic_flow.
	Each sample corresponds to one decision timestamp, and labels are stored
	by index reference to avoid expanding all sensor-by-time targets into labels.csv.

	# Whether to inherit the general processing rules in ../general_task.yaml.
	inherit_general: true

	sample:
	primary_key: timestamp

	# Label storage mode.
	# index_reference: labels.csv does not expand every (ID, timestamp, horizon) label;
	# instead, reference columns point to the underlying time series file.
	# long_table: labels.csv stores each target value in a separate row; this is larger but easier to inspect.
	label_mode: index_reference

	inputs:
	# ------------------------------------------------------------
	# Model input 1: historical traffic flow.
	# ------------------------------------------------------------
	time_series:
	- fields: traffic_flow # Historical time series fields available to the model.
	source: traffic_flow_ca_2019 # Refers to the underlying LargeST h5 file defined in data_refs.time_series.

	window:
	length: 64 # The history window contains 672 task time steps.
	unit: 15min # Each step is 15 minutes, so the model sees the past 7 days.

	sampling: # Optional configuration; set this here if the raw data is not sampled every 15min.
	native_rate: 15min # Native sampling resolution of the LargeST h5 file.
	task_rate: 15min # Input resolution for this task, matching the native resolution.
	downsample_method: mean # If changed to 1h, use mean to aggregate four 15min observations.
	upsample_method: ffill # If changed to a finer resolution, use forward filling.



	tabular_static:
	source: largeST_static_metadata # Refers to the static attribute table in data_refs.tabular_static.

	fields:
	- Lat
	- Lng
	- District
	- County
	- Fwy
	- Lanes
	- Type
	- Direction


	graph:
	source: road_distance_graph

	prediction:
	field: traffic_flow # Target variable to predict.
	source: traffic_flow_ca_2019 # Target values come from the same LargeST historical traffic flow h5 file.

	horizon:
	length: 16 # Predict the next 16 time steps.
	unit: 15min # Each step is 15 minutes, so the forecast covers the next 4 hours.
	starts_after: timestamp # The future window starts after sample.timestamp.

	stride: 16 # Timestamps of adjacent samples are 16 task time steps apart.

	label_storage:
	ref_column: target_series_ref # Reference column in labels.csv pointing to the underlying target sequence.
	shape_column: target_tensor_shape # Column in labels.csv recording the target tensor shape, for example [num_nodes, horizon, num_fields].

	leakage:

	observed_past_only_columns:
	- traffic_flow # traffic_flow can only be a historical input; future traffic_flow is the prediction target and must not be an input.



	data_refs:
	time_series:
	- traffic_flow_ca_2019:
	file: All_Datasets/LargeST_ca_full/input_multimodal/time_series/traffic_flow/ca_his_2019.h5 # LargeST California 2019 historical traffic flow file.
	value_field: traffic_flow # Semantics of the corresponding variable in the h5 file.
	native_grid_minutes: 15 # The native time grid is 15 minutes.



	tabular_static:
	largeST_static_metadata:
	file: All_Datasets/LargeST_ca_full/input_multimodal/ca_meta.csv # Static sensor attribute table.
	id_column: ID

	graph:
	road_distance_graph:
	file: All_Datasets/LargeST_ca_full/input_multimodal/node_level_task_graphs/road_distance_graph/edge_infos.csv