A library for composable forecasting models built on top of scikit-learn
This software library is the open source package for the work done for SETO project 2243, Improving Grid Awareness by Empowering Utilities with Machine Learning and Artificial Intelligence, specifically task 2: Endpoint data process and Analysis. The package includes meter level forecasting and gap filling for customer load as well as photovoltaic (PV) generation forecast models for short term and intra-day PV forecasting.
The package provides the tools to construct machine learning models that fill gaps or forecast in the verification datasets. This includes loading the data and applying the XGBoost estimator as described in section 3.1 of the Final Technical Review (FTR). The composable model framework described below provides the configurable model inputs required for the Neighbor Informed Estimates and Community Analytics. The PV System model in section 3.2 of the FTR is also implemented using the composable model framework. Example PV model construction is shown below.
The API for fit, predict, and metrics is reduced to specifying a start and end times for a given location. The model must construct feature data using column transforms. Having done so, forecasting as a service become trivial.
Users can verify the code works with the example_forecast.py
script by running docker compose up --build
The log output from the container will include feature and weather data as well as predicted values. The script
takes several minutes to run because the weather data is large.
The dockerfile included in the project will run the example_forecast.py which demonstrates both the machine learning model for AMI meter forecasting and the physics based PV model using PySam. Users can then choose between their local working environment and containerized environment to extend and experiment with the time series models library.
The dockerized example is a great place to start for experimentation and further development.
Once the container layers are running (this will take several minutes) the console log will show
[+] Running 1/1
✔ Container seto_forecasting-server-1 Recreated 0.7s
Attaching to seto_forecasting-server-1
The first example is a meter level xgboost estimator
seto_forecasting-server-1 | INFO:__main__:Starting forecast example for AMI meter forecast with XgBoost estimator!
seto_forecasting-server-1 | INFO:time_series_models.transformers:Constructing overfetched range pipeline using lags [ 0 168]
seto_forecasting-server-1 | INFO:time_series_models.processes:Instantiating RegularTimeSeriesModels with kwargs {'day_of_week': True, 'harmonics': array([ 24, 168, 8760], dtype='timedelta64[h]'), 'met_vars': ['t', 'r2'], 'met_horizon': 12, 'mapping': {'p2ulv18716': {'latitude': 35.0, 'longitude': -75.0}}}
There will be a good deal more log messages (and a few minutes to download weather data) for training and prediction
followed by the (truncated) results of the prediction run for the p2ulv18716
meter.
seto_forecasting-server-1 | INFO:time_series_models.data_fetchers.fetcher:Finished 'HrrrFetcher' 'get_data' in 248.3485 secs
seto_forecasting-server-1 | INFO:time_series_models.data_fetchers.fetcher:Finished 'AmiFetcher' 'get_data' in 0.0837 secs
seto_forecasting-server-1 | INFO:__main__:Predicted: predicted true
seto_forecasting-server-1 | location date_time
seto_forecasting-server-1 | p2ulv18716 2021-01-01 00:00:00 785.116394 600.156056
seto_forecasting-server-1 | 2021-01-01 01:00:00 717.555481 2579.214714
seto_forecasting-server-1 | 2021-01-01 02:00:00 817.579041 2720.881345
seto_forecasting-server-1 | 2021-01-01 03:00:00 507.064819 2341.922617
seto_forecasting-server-1 | 2021-01-01 04:00:00 444.800018 2124.941260
seto_forecasting-server-1 | ... ... ...
seto_forecasting-server-1 | 2021-02-04 20:00:00 513.341370 425.300693
seto_forecasting-server-1 | 2021-02-04 21:00:00 591.890686 459.267964
seto_forecasting-server-1 | 2021-02-04 22:00:00 2320.842773 546.954250
seto_forecasting-server-1 | 2021-02-04 23:00:00 2011.579346 599.035650
seto_forecasting-server-1 | 2021-02-05 00:00:00 2223.689941 652.819004
seto_forecasting-server-1 |
seto_forecasting-server-1 | [841 rows x 2 columns]
The PV physics model will continue immediately with
seto_forecasting-server-1 | INFO:__main__:Starting forecast example for PV physical forecast!
Intermediate weather data will be visible during the training and prediction steps, extracted for the time range, latitude and longitude of the site.
seto_forecasting-server-1 | INFO:time_series_models.transformers_pv:Feature DF: ghi dni dhi temp_air wind_speed
seto_forecasting-server-1 | date_time
seto_forecasting-server-1 | 2021-02-01 00:00:00 0.0 0.0 0.0 275.616852 2.302358
seto_forecasting-server-1 | 2021-02-01 01:00:00 0.0 0.0 0.0 273.182281 3.805551
seto_forecasting-server-1 | 2021-02-01 02:00:00 0.0 0.0 0.0 271.327850 2.718800
seto_forecasting-server-1 | 2021-02-01 03:00:00 0.0 0.0 0.0 270.273697 4.061844
seto_forecasting-server-1 | 2021-02-01 04:00:00 0.0 0.0 0.0 270.467178 4.645654
seto_forecasting-server-1 | ... ... ... ... ... ...
seto_forecasting-server-1 | 2021-02-04 20:00:00 586.1 969.0 77.0 281.806671 12.966594
seto_forecasting-server-1 | 2021-02-04 21:00:00 485.4 924.0 72.1 282.243988 15.832714
seto_forecasting-server-1 | 2021-02-04 22:00:00 325.8 547.0 147.3 281.664047 16.190084
seto_forecasting-server-1 | 2021-02-04 23:00:00 128.7 236.0 88.2 279.410828 14.847732
seto_forecasting-server-1 | 2021-02-05 00:00:00 0.0 0.0 0.0 276.510498 12.012517
Finally, the predicted values for the capybara
PV demo site will be logged.
seto_forecasting-server-1 | INFO:time_series_models.transformers_pv:Trying to load from: /app/pv_site.json
seto_forecasting-server-1 | INFO:__main__:pv predictions: predicted
seto_forecasting-server-1 | location date_time
seto_forecasting-server-1 | capybara 2021-02-01 00:00:00 2.555532e+02
seto_forecasting-server-1 | 2021-02-01 01:00:00 2.555532e+02
seto_forecasting-server-1 | 2021-02-01 02:00:00 2.555532e+02
seto_forecasting-server-1 | 2021-02-01 03:00:00 2.555532e+02
seto_forecasting-server-1 | 2021-02-01 04:00:00 2.555532e+02
seto_forecasting-server-1 | ... ...
seto_forecasting-server-1 | 2021-02-04 20:00:00 -1.007929e+06
seto_forecasting-server-1 | 2021-02-04 21:00:00 -1.096734e+06
seto_forecasting-server-1 | 2021-02-04 22:00:00 -9.101299e+05
seto_forecasting-server-1 | 2021-02-04 23:00:00 -4.854458e+05
seto_forecasting-server-1 | 2021-02-05 00:00:00 2.555532e+02
seto_forecasting-server-1 |
seto_forecasting-server-1 | [97 rows x 1 columns]
seto_forecasting-server-1 | INFO:root:All done!
Some systems may return an error code on exit.
seto_forecasting-server-1 | free(): invalid pointer
seto_forecasting-server-1 | Aborted
seto_forecasting-server-1 exited with code 134
This is likely related to an issue with the version of the pysam library but it does not affect the execution of the example.
This library is designed for use by technical engineers and data scientists. It takes advantage of the Python data science ecosystem and therefore requires installation of many third party open source libraries. It has been developed and tested in a Linux operating system. Running on a Docker container such as the canonical Ubuntu image is strongly recommended. The library was developed using Ubuntu 22.04 (Jammy) with Python 3.10.6.
The docker file included in the repository only runs the example_forecast.py
file. The following code snippets
are provided as helpful steps toward building a developer environment where you can run unit tests, forecasting scripts
and jupyter notebooks.
After installing Docker, run the following command to setup a basic Jammy container with this library:
docker run -it ubuntu:jammy /bin/bash
apt-get update
apt-get install git
git clone https://github.com/SETO2243/forecasting.git
cd /forecasting
Reading the grib weather data requires the eccodes library, which is available from the Jammy package repo, installable via shell:
apt-get install libeccodes-dev
Run the following command to install Python package dependencies:
pip3 install -r requirements.txt
Verify that your environment is fully functional by running the automated unit tests:
python -m unittest
This will print "SUCCESS" near the end if the code work correctly in your new environment.
To start jupyter notebook run:
jupyter notebook --NotebookApp.ip=0.0.0.0
This will print a URL, which you can open in your browser. Then open the example notebook and execute the cells in the demonstration to get acquainted with the functionality.
Models can be composed of mixins for various estimators and forecast processes. These composable pieces can be put together in different ways to solve many problems. The RegularTimeSeriesModel is the core that problem specific parts are added to when forecasting or gap filling a particular timeseries. The estimator is the next essential building block. The estimator can be either a Classifier (a discrete estimator) or a Regressor (a continuous estimator). There are many different numerical techniques for supervised learning estimators. The process is the last essential component. It defines the timeseries being forecast and the available feature data that might have predictive value. Having composed a Model class from these three parts, it is then up to the user to create an instance of the class with configuration arguments that tune the model features for the specific meter load or pv forecast.
New models are defined as Python classes, which utilize building blocks provided by this library as base classes. For example, here is the beginning of a model using an Ordinary Least Squares estimator to forecast Balancing Area Hourly Load:
class OlsModel(BalancingAreaHourly, LinearRegressor, RegularTimeSeriesModel):
...
And this example is model using an XgBoost estimator to forecast AMI (smart meter) Hourly Load:
class XgbModel(AmiHourlyForecast, XgbRegressor, RegularTimeSeriesModel):
pass
Additional behaviors including filters and hyperparameter tuning can also be added via composition. See the notebook demo for examples.
The initialization arguments control the model input feature. The configuration arguments are specific to the process and the estimator used to compose the model. The process class (e.g. AmiHourlyForecast) implements the set of possible input features that can be configured. When the model instance is created the specific set of input features for the specific model is defined. Additional parameters can also be passed to the estimator or other components.
As an example, we can configure XgbModel from above with three types of features: lagged features as described in the FTR (Section 3.1.2.2.1), one hot encoded day-of-week values, and a harmonic feature that decomposes time into a sine and cosine waves with the specified frequencies.
config = dict(
lags = np.array([24, 48, 168], dtype="timedelta64[h]"),
day_of_week=True,
harmonics = np.array([24, 168, 365*24], dtype="timedelta64[h]")
)
instance = XgbModel(**config)
Below is the sklearn model diagram for the pipeline of transformations that implement the model.
*Replace this placeholder diagram generated using the camus internal version of the forecast library.
Once a model instance is created, we must train the machine learning algorithm. The fit method takes a start date, stop date and a list of one or more identifiers. For process classes that support it, passing multiple identifiers allows training a single model for a cohort of resources.
instance.fit("2021-01-15", "2021-06-01", "55797646")
Once the model instance is trained, we can call the predict
method (again with a start, stop and one or more
resource identifiers) to generate new predicted values from the input features for the specified date range.
instance.predict_dataframe(
"2021-03-15",
"2021-03-15T04",
"55797646",
range=True,
)
The PV Model as described in the FTR section 3.2 uses the same composable framework define models using the HRRR weather (see below) as an input to the NREL PySam PV generatin algorithm. For the project we used the PySam generation forecast directly using the configration shown below with the IdentityRegressor. Building additional input features for sites with direct telemetry would allow using machine learning models like xgboost too.
pv_config = dict(
lags=None,
site_config_mapping={
"capybara": ["/app/pv_site.json"],
},
site_latlong_mapping={
"capybara": dict(
latitude=40.0,
longitude=-100.0,
),
},
site_meter_mapping=None,
source_mode="12_hour_horizon",
)
class PVForecastModel(
PVForecast,
IdentityRegressor,
RegularTimeSeriesModel,
):
pass
pv_instance = PVForecastModel(**pv_config)
pv_instance.model
*Replace this placeholder diagram generated using the camus internal version of the forecast library.
Engineers and data scientists commonly use an interactive web-based development environment called Jupyter Notebook (now Jupyter Lab) to explore and visualize data and algorithms in a cell based execution environment.
An example notebook is provided in this GitHub repository which demonstrates the core capabilities of the time series models library developed for the SETO project.
Sample AMI meter data for 4 meters is available for validation in a public GCS bucket: gs://seto2243-forecasting/ami_validation Customer data used in the project is available upon completion of an NDA as described in the FTR.
The High Resolution Rapid Refresh (HRRR) forecast is an operational weather forecasting product of the NOAA Center for Environmental Prediction. The HRRR forecast results are publicly available on multiple cloud vendor platforms (AWS, GCP, Azure) via a public private partnership with the NOAA Open Data Dissemination program.
These cloud providers host the petabyte scale archive of grib2 files created by the hourly HRRR operational forecast system. As part of the project a set of metadata file that index the archive was created using an open source tool called Kerchunk. The weather data is a public archive provided by NODD. To use it the timeseries models developed for the project a kerchunk metadata index must be created. A sample index is provided in the project GCS bucket.
A kerchunk aggregation of select variables from the 12 hour forecast for 2021 is available in a public GCS bucket: gs://seto2243-forecasting/high-resolution-rapid-refresh
Library code is organized into a number of subpackages described below, to aid engineers working on writing or debugging code using this library:
data_fetchers/
: data fetchers and associated tests -- see directory-specific README.back_test.py
: BackTest class for estimating model performance on out-of-sample data using a time series (rolling) cross-validation.config.py
: ConfigHandler helper class to encode and decode time series models configurationsconstants.py
: defines constant variables referenced throughout Time Series Models librarydata_monitor.py
: ForecastDataMonitor is a pipeline element that collects statistics on the data passing through itdecorators.py
: functions that can be used as decorators in Time Series Models librarydummy_models.py
: classes implementingDummyDataModel
for using external predictions with Time Series Models metrics and vizualizationsestimators.py
: Estimator and EstimatorWrapper mixins for forecast models (e.g.,XgbRegressor
,RandomizedSearch
)filters.py
: Filter mixins (e.g.,DropFeatures
,RowFilter
,StandardScaler
)- Gap solving tools:
gap_maker.py
: utility for introducing synthetic gaps into a datasetgap_filler.py
: utility for filling data gaps with various interpolative or extrapolative methodsgap_runner.py
: holds score_df_at_locs, a helper to evaluate gap-filling performance
metrics_runner.py
: a container for running models and comparing metrics, organized around monthly analysisprocesses.py
: mixins defining the data sources and preprocessing steps for a given process to forecastpv_physical_model.py
: common interface to using pvlib and PySAM physical models- Visualization tools:
shap_viz.py
: wrapper for applyingshap
library to RegularTimeSeriesModel instancesviz.py
: methods for plotting time series of predicted vs actual, and residual scatter plots
time_series_models.py
: coreRegularTimeSeriesModel
andMixin
class definitions, along with several other model helpers and mixins likeAggModel
andMonthSelector
.time_unit.py
: Enum for calendar conversion of numpy datetime64 arrays- Transformers: a collection of transformers for composing model features
transformers.py
: broad-purpose (or uncategorized) transformers for use throughout Time Series Models- core helper functions like
make_domain
- some bespoke transformers that haven't been split out into separate modules
- simple array transformations
- core helper functions like
transformers_calendar.py
: calendar feature transformers (e.g., business day, day-of-week, & harmonics)transformers_pv.py
: transformers for PV forecastingtransformers_shortcuts.py
: help avoid copy-pasting the same core features for each process (forecast sum of meters & DER, rather than sum of forecasts)
version.py
: Time Series Models library version