7.7.3. Generic Time Series Forecasting

This example serves as a “Hello World” introduction to time series forecasting using the TinyML ModelMaker toolchain. It demonstrates how to use any generic time series forecasting task with our toolchain.

7.7.3.1. Overview

Task: Time series forecasting (predict future values)
Dataset: Simulated thermostat data (temperature oscillation)
Model: FCST_LSTM10 (~542 parameters)
Target: F28P55 (NPU device)

7.7.3.2. Running the Example

cd tinyml-modelzoo
./run_tinyml_modelzoo.sh examples/generic_timeseries_forecasting/config.yaml

cd tinyml-modelzoo
run_tinyml_modelzoo.bat examples\generic_timeseries_forecasting\config.yaml

7.7.3.3. Understanding the Dataset

The dataset models a room temperature controlled by an ON/OFF heater with hysteresis:

Heater turns ON when temperature drops below 20C
Heater turns OFF when temperature rises above 24C
Temperature changes gradually due to thermal inertia

Column	Description
`timestamp`	Time index
`temperature`	Room temperature (C)

Dataset split:

Training: 10 files (thermostat_01.csv to thermostat_10.csv)
Validation: 2 files
Test: 3 files

Each file contains 1000 timesteps of simulated temperature data.

7.7.3.4. Dataset Format

For forecasting tasks, ModelMaker expects the dataset in this structure:

{dataset_name}.zip/
|
|-- files/
|     |-- {file1}.csv
|     |-- {file2}.csv
|     |-- {fileN}.csv
|
|-- annotations/
      |-- file_list.txt
      |-- instances_train_list.txt
      |-- instances_val_list.txt
      |-- instances_test_list.txt

Note

Unlike classification tasks, forecasting always requires annotation files to specify train/validation/test splits.

7.7.3.5. Configuration

common section:

common:
  task_type: generic_timeseries_forecasting
  target_device: F28P55

dataset section:

dataset:
  dataset_name: generic_timeseries_forecasting
  input_data_path: https://software-dl.ti.com/C2000/esd/mcu_ai/01_03_00/datasets/generic_timeseries_forecasting.zip

data_processing_feature_extraction section:

data_processing_feature_extraction:
  data_proc_transforms:
    - SimpleWindow  # Mandatory for forecasting
  frame_size: 32
  stride_size: 0.1
  forecast_horizon: 2  # Predict 2 future timesteps
  variables: 1
  target_variables:
    - 0  # Column index (temperature)

training section:

training:
  model_name: FCST_LSTM10
  batch_size: 32
  training_epochs: 50
  num_gpus: 1
  quantization: 1
  optimizer: adam
  output_int: false  # Must be false for forecasting

Important

output_int must be set to false for forecasting tasks.

7.7.3.6. Target Variables

The target_variables parameter can be specified in multiple formats:

Format	Example	Description
Empty list	`[]`	Predict all columns
List of indices	`[0]` or `[0, 1, 2]`	Column indices (0-indexed)
List of names	`['temperature']`	Column names from header
Single index	`0`	Single column index
Single name	`temperature`	Single column name

7.7.3.7. Evaluation Metrics

SMAPE (Symmetric Mean Absolute Percentage Error)

Measures percentage error between predicted and actual values.

Range: 0% to 200%
Ideal value: 0% (lower is better)

R2 Score (Coefficient of Determination)

Indicates how well predictions match actual values.

Range: (-infinity, 1]
Ideal value: 1 (higher is better)

7.7.3.8. Expected Results

Float Training (Best Epoch 48):

Variable	Timestep 1	Timestep 2	Overall
SMAPE	0.60%	0.88%	0.74%
R2	0.9811	0.9563	0.9687

Test Results:

Variable	Timestep 1	Timestep 2	Overall
SMAPE	0.66%	0.95%	0.80%
R2	0.9763	0.9517	0.9640

7.7.3.9. Output Location

Results are saved to:

../tinyml-modelmaker/data/projects/{dataset_name}/run/{date-time}/{model_name}/

Key outputs:

training/base/best_epoch_{N}_results/ - Float training results
training/quantization/best_epoch_{N}_results/ - Quantized results
training/quantization/test_results/ - Test results with prediction plots
compilation/artifacts/mod.a - Compiled for device

7.7.3.10. Next Steps

After successfully running this generic example:

Try PMSM Rotor Forecasting - Real-world motor temperature forecasting
Try HVAC Indoor Temp Forecast - HVAC temperature prediction
Explore Forecasting Dataset Format - Use your own forecasting data
Read Understanding Config - Learn all config options