Videos

OpenSTEF (Open Short-Term Energy Forecasting) is an open source Python package that provides automated machine learning pipelines for short-term energy load forecasting.
OpenSTEF provides automated machine learning pipelines that forecast energy load at specific locations up to two weeks ahead.

OpenSTEF is collaboratively developed by a growing community of grid operators, technology vendors, researchers, and energy-sector experts, and is hosted at LF Energy. It generates probabilistic forecasts for hours to days ahead for any given energy signal, enabling grid operators, energy companies, and researchers to anticipate congestion, support grid safety analysis, and optimize flexible assets. The latest release, OpenSTEF 4.0, introduces a fully redesigned modular architecture that broadens the range of forecasting applications and reduces implementation effort.

Learn more at https://openstef.energy.

2 0

This technical working group meeting covers progress on CalTRACK development as the open-source libraries transition to Open DSM under the Linux Foundation. The team introduces Open DSM's new logo and outlines ongoing work to complete online documentation for the daily, billing, and hourly models. Additionally, they present a comprehensive overview of the new CalTRACK 3.0 hourly model under the hood, detailing its elastic net framework, key inputs—such as temperature and global horizontal irradiance (GHI)—and its autocorrelation imputation methods.1

The group also examines temporal clustering and interaction-based temperature binning to prevent categorical and extreme-temperature overfitting. Finally, the speakers share population testing results across 3,500 residential and commercial solar and non-solar meters, showing significant CVRMSSE improvements and bias reductions, and outline a roadmap for model testing and code release in early 2025.

0 0

This technical working group meeting covers progress on the CalTRACK 3.0 hourly model alongside the release of updated daily and billing documentation live on the OpenEEmeter repository. The team introduces two critical model modifications to resolve categorical and temperature-based overfitting: temporal clustering—which dynamically groups 84 month-day categories into 2 to 24 highly representative behavioral clusters—and interaction-based temperature binning to generate unique slopes and intercepts.1

Additionally, Recurve presents population-level cross-validation results across 35,000 residential and commercial meters, evaluating joint candidate models against legacy CalTRACK 2.0 (OpenEEmeter 3.0). While commercial and extreme temperature bins present minor trade-offs, the unified candidate framework handles solar and non-solar customers consistently, delivering massive error reductions and resolving underprediction bias for solar PV customers on cloudy days while operating two to three times faster than the legacy model.

0 0

This technical working group meeting covers development progress on the CalTRACK 3.0 hourly model and introduces new OpenEEmeter documentation. The team reviews updated online documentation for the daily and billing models and highlights the upcoming guide for the hourly model. They share performance testing comparing a 24-hour day-based prediction model to a per-hour prediction model, demonstrating that the day-based approach achieves superior predictive accuracy (PNRMSSE) and runs significantly faster.1

Additionally, Recurve data scientists address temperature-related model bias by introducing a new temperature-binning feature that appends one-hot encoded temperature vectors to the model coefficients. To solve overfitting caused by having too many time-based categories (such as 84 combinations of weekdays and months), the team introduces a new daily usage clustering module. This module utilizes silhouette values to dynamically reduce the 84 categories down to 2–4 custom clusters per meter, capturing distinct building behaviors (like summer cooling peaks and weekend-weekday shifts) while drastically reducing model noise. Finally, they outline an updated hyperparameter optimization process that now incorporates both PNRMSSE and seasonal bias (PNMBE) in its objective function to finalize temperature bins, alpha, and percent L1 parameters.

0 0

This technical working group meeting covers final-stretch development progress on the CalTRACK 3.0 hourly model. The team introduces a multi-variate radial basis function (RBF) interpolation procedure to fill in missing 24-hour data points and prevent model crashes, sharing details on its accuracy and performance. They discuss the flexibility of the hourly model framework in incorporating supplemental time-series data, specifically examining how using a solar PV installation date vector helps model buildings with partial baseline solar data. Additionally, the team presents successful hyperparameter optimization results for the solar and non-solar elastic net models, demonstrating stable, well-behaved response surfaces for alpha and percent lasso variables. Finally, they share population testing results highlighting significant error and overfitting improvements over CalTRACK 2.0, particularly for solar PV customers on cloudy days.

1 0

This technical working group meeting covers development progress on the CalTRACK 3.0 hourly model and introduces library integration proposals. The team discusses a proposal to donate the "grid meter" library to OpenEEmeter to introduce comparison groups that correct model predictions using population trends and external factors. They also explore renaming OpenEEmeter to a generic placeholder to act as a unified monorepo umbrella for EEmeter, grid meter, and EEweather. Additionally, Recurve data scientists review their transition from a multivariate radial basis function (RBF) interpolator to a highly flexible autocorrelation interpolation method to resolve missing temperature and observed data points. They share comprehensive testing results of this method across more than 8,000 residential and commercial meters. Finally, the team presents population-level bias results for the current daily and hourly models, showing that while predictive error is minimized, a temperature-related bias exists. To solve this, they propose a dynamic temperature-binning approach to linearize response curves under the hood of the model.

0 0

This technical working group meeting covers progress on the new CalTRACK 3.0 hourly model and its integration into the OpenEEmeter codebase. The team reviews the transition of code from R&D to the OpenEEmeter API, demonstrating a highly compatible fit-and-predict architecture. Additionally, they present results from a hyperparameter optimization study analyzing the elastic net's L1 and L2 normalization parameters across 36,000 solar and non-solar models. The findings confirm the robustness of the global optimal hyperparameters and highlight the importance of the elastic net over simple linear regression in preventing overfitting, while also demonstrating how to identify and handle anomalous meter data.

0 0

This technical working group meeting covers key progress on CalTRACK development and features updates on the OpenEEmeter 4.0 release. The team announces the public release of OpenEEmeter 4.0 on PyPI, alongside an upcoming Linux Foundation press release and webinar. Additionally, Recurve data scientists provide an in-depth review of R&D efforts for the CalTRACK 3.0 hourly model. They share comparative performance testing between an LSTM-based deep neural network and an elastic net regressor, examining the tradeoffs in computation time, data limitations, and overfitting. The team also demonstrates the flexibility of the hourly model framework in incorporating supplemental time-series data (such as EV charging and pumping schedules) and reviews updated solar meter test results using a revised cloudiness percentage index.

0 0

This technical working group meeting covers progress on integrating daily, billing, and hourly models into OpenEEmeter 4.0. The team details how OpenEEmeter 4.0 delivers the daily model's seasonal and weekday/weekend bias reductions while maintaining a legacy 3.0 mode (CalTRACK 2.0 equivalent) with matching results. They share plans to combine OpenEEmeter and EEweather into a single library to unify shared functions and streamline temperature data flow. Additionally, the team introduces an updated API architecture inspired by scikit-learn's fit/predict syntax to improve user and developer ease of use, and details how the data class will propagate data-sufficiency and model-fit violations.1

The group also votes to pass forward a data sufficiency change: allowing negative values for daily electric meters to accommodate overproducing solar customers, while restricting negative values exclusively for non-electric meters. Finally, Recurve presents a comprehensive cross-validation analysis of hourly regression models on a three-year rolling test-train scheme across 3,000 residential solar and non-solar meters. Test results show that an elastic net framework utilizing temperature and global horizontal irradiance (GHI) significantly reduces test error, mitigates overfitting, improves computational efficiency, and drastically minimizes underprediction bias on cloudy days compared to the legacy model.

0 0