Trace Prediction

This example turns a recorded thermostat trace into a supervised prediction problem. It loads the CSV with DataFrame.from_csv, preserves time order with TrainTestSplit(shuffle=False), applies SlidingWindow(window_size=4), trains a RegressionTree with learn_offline, and evaluates the held-out windows with evaluate_offline.

Run it from the repository root:

uv run --directory ./examples/trace_prediction python run.py

The output starts with the data flow and then prints the model results:

Trace prediction example
source trace rows: ...
transform: SlidingWindow(window_size=4)
training windows: ...
test windows: ...
evaluation:
prediction preview:

Regenerate the deterministic trace data with:

uv run --directory ./examples/trace_prediction python generate_data.py

Reusable notebook snippet:

from pathlib import Path

from flowcean.core import learn_offline
from flowcean.polars import DataFrame, SlidingWindow, TrainTestSplit
from flowcean.sklearn import RegressionTree

inputs = [
    f"{column}_{step}"
    for step in range(3)
    for column in ("temperature", "target", "heating")
]
outputs = ["temperature_3"]

trace = DataFrame.from_csv(Path("examples/trace_prediction/data/thermostat_trace.csv"))
train, test = TrainTestSplit(ratio=0.8, shuffle=False).split(trace)
window = SlidingWindow(window_size=4)

model = learn_offline(
    train | window,
    RegressionTree(max_depth=5, random_state=42),
    inputs,
    outputs,
)

Key Flowcean pieces:

• DataFrame wraps the lazy Polars CSV data as an offline environment.
• TrainTestSplit(shuffle=False) creates train and test environments without disrupting trace order.
• SlidingWindow(window_size=4) builds supervised rows from sequential samples.
• learn_offline trains the learner on the training environment.
• evaluate_offline computes regression metrics on the test environment in the full script.