LF Energy OpenDSM Completes Development of New Hourly Model
The LF Energy OpenDSM project is excited to announce the completion of a new hourly modeling framework for OpenEEmeter, designed to deliver faster, more accurate, and more flexible energy consumption predictions, especially for customers with solar PV.
This major update was driven by a set of clear goals:
- Improve solar PV predictions by integrating solar irradiance data
- Match or exceed the accuracy of the prior model for non-solar predictions
- Reduce overfitting so baseline performance better reflects prediction accuracy
- Support arbitrary time-series inputs for research and development
- Run significantly faster to enable large-scale deployments
We’re pleased to report that the new model meets, and in many cases exceeds, these goals.
Key Improvements
- Solar accuracy de-risked: The new solar model integrates Global Horizontal Irradiance (GHI) data to handle variability from cloud cover, reducing the risk of poor predictions for solar PV customers (use of the non-solar model does not require GHI).
- Comparable non-solar performance: Accuracy for non-solar meters is on par with the legacy model while reducing overfitting, making error metrics more trustworthy.
- Less overfit: Baseline and reporting-period errors now align much more closely, improving confidence in predictions.
- Adaptive outlier down-weighting: Outliers in the training data are automatically downweighted to prevent them from adversely affecting the overall fit.
- R&D-ready: The framework supports additional arbitrary time-series or categorical variables for research use.
- 4-5x faster: The new approach cuts processing time dramatically, reducing costs for large-scale runs.
Development Process
This work was reviewed monthly with feedback from the OpenDSM working group, regularly drawing 15-20 active participants. Development and validation used a dataset of 33,000 meters across California, covering residential and commercial customers with and without solar PV.
The model’s Elastic Net regression framework balances predictive power with computational efficiency, while its 24-hour input-output architecture captures correlations across the entire day. A robust, adaptive loss function helps manage outliers without distorting fit quality.
Learn More
For the full technical details – including methodology, hyperparameter optimization, and population-level results – read the comprehensive Hourly Model Report here.
Also, join us for a free webinar on September 24 at 9am US Pacific to learn more about the project. Register at https://community.linuxfoundation.org/events/details/lfhq-lf-energy-presents-reliably-measuring-demand-side-interventions-in-a-solar-pv-world-with-opendsm/.