7.9.29. Forecasting Examples Guide

In 1.4.0 this page was a generic forecasting tutorial that pointed to an examples/forecasting folder. That folder does not exist in 1.5.0. This page is now a short guide to the forecasting examples and the settings that matter for them. Each example has its own page with the full configuration.

7.9.29.1. Forecasting Examples in 1.5.0

All use task_type: generic_timeseries_forecasting and are run from the tinyml-modelzoo folder with ./run_tinyml_modelzoo.sh examples/<folder>/<config>.yaml (Windows: see Running Examples).

Config	Device	Model	Page
`generic_timeseries_forecasting/config.yaml`	F28P55	`FCST_LSTM10`	Generic Timeseries Forecasting
`forecasting_pmsm_rotor_temp/config.yaml`	F28P55	`FCST_LSTM8`	PMSM Rotor Temperature Forecasting
`forecasting_pmsm_rotor_temp/config_MSPM0.yaml`	MSPM0G5187	`FCST_LSTM8`	PMSM Rotor Temperature Forecasting
`hvac_indoor_temp_forecast/config.yaml`	F28P55	`FCST_LSTM10`	HVAC Indoor Temperature Forecasting

7.9.29.2. How Forecasting Works

A forecasting model reads a window of past values (frame_size time steps of variables input columns) and predicts the next forecast_horizon steps of one or more target_variables. The SimpleWindow transform creates the windows and is mandatory. The stride (stride_size) is a fraction of the frame size.

data_processing_feature_extraction:
  data_proc_transforms:
  - SimpleWindow
  frame_size: 32
  stride_size: 0.1
  forecast_horizon: 2
  variables: 1
  target_variables:
  - 0

Rules to remember:

The dataset needs the files/ plus annotations/ layout, see Forecasting Dataset Format.
output_int must be false in the training section.
Feature extraction is not supported for forecasting. Downsample can be added to data_proc_transforms together with sampling_rate and new_sr.
target_variables accepts column indices (0-based after the time column), column names, a single index or name, or an empty list to predict all columns.
Multi-variable forecasting works by listing several indices or names in target_variables. Metrics are reported per variable and per predicted time step.

7.9.29.3. Metrics

SMAPE (0% to 200%, lower is better) and R2 (higher is better) are computed per target variable and per predicted time step. The best epoch is the one with the lowest average SMAPE. A larger forecast_horizon normally lowers the accuracy of the later time steps, which you can see in the per-time-step table of each run.

7.9.29.4. Available Models

The forecasting models enabled in tinyml_modelzoo/model_descriptions/forecasting.py are:

Model	Description
`FCST_LSTM8`, `FCST_LSTM10`	One LSTM layer (hidden size 8 or 10) followed by a linear layer
`FCST_3k`	4 fully connected layers, about 3k parameters
`FCST_13k`	2 Conv, BatchNorm, ReLU layers and a linear layer, about 13k parameters
`FCST_500_NPU`, `FCST_1k_NPU`, `FCST_2k_NPU`, `FCST_4k_NPU`, `FCST_6k_NPU`, `FCST_8k_NPU`, `FCST_10k_NPU`, `FCST_20k_NPU`	NPU-compliant convolutional models of about 0.5k to 20k parameters (`FCST_6k_NPU` uses depthwise separable convolutions)

Use an _NPU model on devices with an NPU (F28P55, F28P65, MSPM0G5187) and see Model Zoo Reference.

Warning

The quantized FCST_LSTM10 model has shown a large accuracy drop in 1.5.0 runs (see Generic Timeseries Forecasting). Always compare the float and quantized results of your own run.

7.9.29.5. Forecasting versus Regression

Aspect	Regression	Forecasting
Output	Value for the current window	Future values (`forecast_horizon` steps)
Task type	`generic_timeseries_regression`	`generic_timeseries_forecasting`
Metrics	RMSE, R2	SMAPE, R2 per time step
Examples	Generic Timeseries Regression, Torque Measurement Regression	Generic Timeseries Forecasting, PMSM Rotor Temperature Forecasting

7.9.29.6. Deployment

See Deploying Forecasting Models from ModelMaker to Device.