7.9.28. HVAC Indoor Temperature Forecasting

Forecast the indoor temperature of a room from compressor frequency and outdoor temperature, as an input to predictive HVAC control.

Example folder: examples/hvac_indoor_temp_forecast (GitHub)

7.9.28.1. Overview

Traditional HVAC controllers react to temperature deviations after they occur, which causes delayed corrections, unnecessary compressor activity and higher energy use. A forecast of the indoor temperature lets a higher-level controller plan smoother, more energy-efficient compressor actions. This example trains a small LSTM offline and compiles it to run fully on the device.

Task: timeseries forecasting (generic_timeseries_forecasting)
Device: F28P55 (config.yaml), compiled with the forced_soft_npu_preset (software emulation of the NPU)
Model: FCST_LSTM10
Target: indoorTemperature at the next time step (set forecast_horizon for a longer horizon)

7.9.28.2. The Dataset

A synthetic dataset that emulates indoor thermal behavior. It has uniformly sampled values of:

compressorFrequency
outdoorTemperature
indoorTemperature

The model input is the past 5 values of each of the three signals (frame_size: 5, variables: 3) and the target is indoorTemperature (column index 2, counting from 0 after the time column). The config downloads https://software-dl.ti.com/C2000/esd/mcu_ai/datasets/hvac_indoor_temp_forecast_dataset.zip. The layout is the forecasting layout described in Forecasting Dataset Format.

Note

In the automated 1.5.0 config test run this config failed while extracting the download (“File is not a zip file”). If you see the same error, check that the dataset URL returns a valid zip before debugging the pipeline.

7.9.28.3. Running the Example

Install Tiny ML ModelZoo first (Installation, Quickstart), then run from the tinyml-modelzoo folder:

./run_tinyml_modelzoo.sh examples/hvac_indoor_temp_forecast/config.yaml

See Running Examples for all launchers, for example:

run_tinyml_modelzoo.bat examples\hvac_indoor_temp_forecast\config.yaml

7.9.28.4. Configuration

common:
  task_type: generic_timeseries_forecasting
  target_device: F28P55
dataset:
  dataset_name: hvac_indoor_temp_forecast
  input_data_path: https://software-dl.ti.com/C2000/esd/mcu_ai/datasets/hvac_indoor_temp_forecast_dataset.zip
data_processing_feature_extraction:
  data_proc_transforms:
  - SimpleWindow
  frame_size: 5
  stride_size: 0.2
  forecast_horizon: 1
  variables: 3
  target_variables:
  - 2
training:
  model_name: FCST_LSTM10
  model_config: ''
  batch_size: 6
  training_epochs: 10
  early_stopping: true
  early_stopping_patience: 5
  num_gpus: 0
  quantization: 1
  optimizer: adamw
  learning_rate: 0.003
  weight_decay: 0.001
  lr_scheduler: none
  output_int: false
testing: {}
compilation:
  compile_preset_name: forced_soft_npu_preset

target_variables also accepts column names or several indices. output_int must be false for forecasting.

7.9.28.5. Results

The example readme reports these results for the indoor temperature (time step 1):

Stage	SMAPE	R2
Float train, best epoch	0.30%	0.9971
Quantized train, best epoch 4	0.80%	0.9893
Test set	0.73%	0.9857

Warning

These numbers come from the example readme and were not reproduced in 1.5.0 because the test run could not download the dataset. Quantized FCST_LSTM10 models have shown a large accuracy drop in other 1.5.0 runs (see Generic Timeseries Forecasting), so compare the float and quantized results of your own run.

Predicted versus actual indoor temperature

7.9.28.6. Output Location

Results are written to ./data/projects/hvac_indoor_temp_forecast/run/<date-time>/FCST_LSTM10/. The device files are compilation/artifacts/mod.a, compilation/artifacts/tvmgen_default.h and the golden vectors in training/quantization/golden_vectors/.

7.9.28.7. Running on Device

Follow Deploying Forecasting Models from ModelMaker to Device. After flashing, the model output matches the golden vectors.

On-device results — On-device output matching the golden vectors

Device measurements from the readme (F28P55x, written for an earlier release):

Metric	Value
AI model cycles	199432
Inference time	1329.55 us
Model size (flash / SRAM)	9174 bytes / 278 bytes
Results match golden vectors	TRUE