Abstract
Integrating renewable energy into local microgrids requires balancing erratic solar production and building demand under significant uncertainty. Historically, the field of Energy Management Systems (EMS) has faced a reproducibility crisis due to the use of private datasets and non-standardized testing environments. This paper introduces EMSx, an open-source benchmarking ecosystem comprising a massive real-world dataset from 70 industrial sites and a standardized Julia-based testing framework. Our evaluation of baseline algorithms demonstrates that Stochastic Dynamic Programming with auto-regressive modeling (SDP-AR) significantly outperforms traditional Model Predictive Control (MPC), effectively closing the gap toward theoretically optimal energy management.
Introduction
As the world shifts toward decentralized energy, microgrids have become essential for integrating local renewable sources like photovoltaics (PV). However, managing these systems is a complex optimization problem. EMS controllers must decide in real-time when to charge or discharge batteries to minimize costs while accounting for highly volatile local demand and generation.
Until now, comparing different control strategies has been nearly impossible. Researchers often validate their algorithms on isolated case studies—such as a single building or a specific transit station—using private data. This lack of transparency prevents the identification of truly robust strategies. EMSx addresses this by providing a unified, transparent, and realistic "sandbox" for microgrid control.
The EMSx Framework
The EMSx benchmark is built on three pillars designed to facilitate fair comparison and rapid iteration:
- Mathematical Framework: A standardized model of a microgrid consisting of one PV unit, one battery, and one load. Performance is measured using a "Performance Score" that scales a controller's savings between a "do-nothing" baseline (0.0) and a theoretical "perfect foresight" upper bound (1.0).
- Software (EMSx.jl): An open-source Julia package optimized for high-performance parallel simulation. It allows researchers to implement and test new controllers against millions of historical data points with minimal code.
- Baseline Algorithms: The framework includes standard industry and academic methods, ranging from deterministic Model Predictive Control (MPC) to advanced "cost-to-go" methods like Stochastic Dynamic Programming (SDP).
Dataset Diversity and Uncertainty
The core of EMSx is a dataset provided by Schneider Electric, featuring 15-minute resolution data from 70 diverse industrial sites across Europe and the United States. This scale is unprecedented in the literature, involving roughly 1.6 million individual control calls per benchmark run.
A key challenge identified in the benchmark is the "Stochastic Gap." Because local demand is far more erratic than national grid demand, deterministic models often fail to account for the long-term value of stored energy. The dataset captures a wide range of predictability across sites, forcing controllers to prove their robustness in both stable and highly volatile environments.
Key Results and Insights
The benchmarking results reveal a clear hierarchy in controller performance. Deterministic MPC, the current industry standard, achieved a performance score of 0.487. While stochastic lookahead methods (OLFC) provided marginal improvements (0.513), the most significant gains came from incorporating "memory" into the state space.
The SDP-AR(1) model, which adds a single lag of net demand to the state space, emerged as the clear winner with a score of 0.794. This approach allows the controller to "sense" current trends in erratic data. Interestingly, increasing complexity further (e.g., moving to AR(2)) increased training time tenfold with negligible performance gains, identifying a practical "sweet spot" for deployment.
Beyond accuracy, SDP-AR(1) demonstrated superior computational efficiency. While it requires an offline training phase (~38 seconds), its online decision time is approximately 0.4 ms—orders of magnitude faster than stochastic lookahead methods, which can take up to 200 ms per decision.
Conclusion
EMSx establishes a new gold standard for evaluating microgrid energy management. By providing a massive, standardized dataset and a high-performance testing framework, it enables the community to move beyond isolated case studies. The success of SDP-AR(1) highlights the importance of state-space design in handling local uncertainty. Future work will expand the benchmark to include multiple storage types and controllable loads, while inviting the community to test Reinforcement Learning and Robust Optimization techniques against this open challenge.
Summary updated