DER Forecast Model Selection Using Cost-Feedback Simulation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing forecasting models for distributed energy resource management systems (DERMS) rely on generic accuracy metrics that do not align with the unique optimization objectives of DERMS, leading to suboptimal energy cost outcomes and failure to adapt to site-specific changes.
Innovation Solution
A method for selecting and training forecasting models that evaluates performance based on actual energy cost outcomes using optimization simulations, incorporating site-specific constraints and penalties for over-prediction errors to align with cost minimization and charging demands.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If generic accuracy metrics (RMSE, MAE, MSE) are used to evaluate forecasting models, then model selection is simplified and computationally efficient, but the models fail to align with DERMS optimization objectives and do not minimize actual energy costs
Solution Approach 1:
The patent implements a feedback mechanism where energy cost results from optimization simulations are fed back into the model training process. The system uses these cost outcomes to adjust model parameters and select models that actually minimize energy costs rather than merely achieving low RMSE/MAE values. This closed-loop feedback ensures models are evaluated and improved based on their real impact on DERMS optimization objectives.
Solution Approach 2:
The system changes the evaluation parameters from generic accuracy metrics (RMSE, MAE, MSE) to domain-specific cost metrics that reflect actual DERMS objectives. By transforming the performance measurement parameters to align with energy cost minimization goals, the system selects models that are truly effective for the intended application rather than those that merely perform well on standard statistical measures.
2Device complexity
If standard forecasting models are deployed without site-specific customization, then deployment time and complexity are reduced, but the models cannot adapt to changing customer operations, pricing structures, or equipment configurations
Solution Approach 1:
The patent implements dynamic model retraining capabilities that allow the forecasting system to adapt to changing site conditions. When customer operations, pricing structures, or equipment configurations change, the system automatically retrains models using updated data and re-evaluates them through optimization simulations. This dynamic adaptation ensures models remain aligned with current DERMS objectives without requiring complete redeployment.
Solution Approach 2:
The system performs preliminary evaluation of forecasting models through optimization simulations before final deployment. By pre-testing models against site-specific constraints, pricing structures, and equipment configurations using simulated operating set points, the system ensures models are properly customized and validated for each unique site before full implementation, balancing customization needs with deployment efficiency.
3Reliability
If optimization simulations with site-specific constraints are run to evaluate model performance, then model selection aligns with energy cost minimization goals, but computational time and processing resources increase significantly
Solution Approach 1:
The patent applies partial action by evaluating only the most promising candidate models through full optimization simulations. Rather than running exhaustive simulations on all possible models, the system first filters candidates using quicker preliminary assessments, then applies computationally intensive simulation-based evaluation only to the top contenders. This selective approach maintains high selection accuracy while reducing overall computational time and resource requirements.
Data Source
AI summary
Certain aspects of the present disclosure provide techniques for selecting and training a forecasting model for a distributed energy resource management system. An example method includes selecting a forecasting model from a plurality of candidate forecasting models; training the selected forecasting model using accuracy measurements comprising at least one of root mean squared error (RMSE), mean absolute error (MAE), or mean squared error (MSE); generating forecasting output data from the trained forecasting model based on predetermined test data; running a simulation using the forecasting output data, the simulation utilizing optimization logic and constraints to simulate control behaviors of distributed energy resources, the simulation being run by generating operating set points based on the forecasting output data to yield an energy cost result; and providing the energy cost result as feedback for determining differential weightings applied to one or more of the accuracy measurements and selection of an updated forecasting model.