TL;DR
This paper introduces Transient Predictive Control (TPC) as a practical, model-free approach for grid-connected inverter control. By demonstrating TPC's ability to run efficiently on standard hardware and effectively manage inverter behavior in real-world grid scenarios, the authors pave the way for more robust and "plug-and-play" integration of renewable energy sources.
What Problem Does This Paper Address?
Integrating renewable energy sources like solar and wind into the power grid via inverters presents a significant control challenge. Traditional methods rely on accurate grid models, which are often difficult to obtain and maintain, especially in dynamic distribution grids. This lack of precise knowledge hinders the development of "plug-and-play" inverter control systems that can seamlessly and reliably connect to the grid. Existing data-driven methods, like Data-Enabled Predictive Control (DeePC), often suffer from high computational demands, slow execution speeds, and issues with causality and bias when using real-world data. The authors aim to bridge this gap by proposing and validating a data-driven control strategy that is both computationally efficient and effective in real-world grid conditions.
Methods
The core of this paper is the application and validation of Transient Predictive Control (TPC). TPC is a model-free predictive control technique that addresses the limitations of traditional methods and other data-driven approaches like DeePC.
Here's a breakdown of the key methodological aspects:
- Model-Free Approach: TPC does not require an explicit mathematical model of the grid or the inverter. Instead, it learns the system's dynamics directly from input-output data.
- Offline Training, Online Optimization: TPC processes historical input-output data offline to estimate a "Multistep Predictor" (denoted by $\hat{H}$). This predictor captures the system's dynamics. The actual control decisions are then made online by solving a computationally tractable optimization problem that uses this pre-computed predictor.
- Transient Predictor Method: The paper refers to Algorithm 1 for estimating $\hat{H}$. This algorithm is designed to produce a consistent and causal estimate of the predictor, even when trained on data collected under closed-loop operation.
- Online Optimization Problem: The online control loop involves solving a Quadratic Program (QP) or Second-Order Cone Program (SOCP) to determine the optimal future control inputs (e.g., current references). This optimization is designed to be fast and independent of the amount of training data used.
- Output Current Constraint: A crucial aspect for inverter control is managing strict hardware current limits. The paper implements an indirect method for constraining the output current magnitude by including the $d$ and $q$ components of the current in the output vector $y$ and formulating the constraint as a Second-Order Cone Program (SOCP). This is preferred over directly constraining the magnitude $|i|$ as it leads to a more LTI-friendly formulation.
Datasets
The paper describes two experimental setups where TPC was tested:
- Two-Converter Experiment: This setup involved two 3 kVA back-to-back DC-AC converters. One acted as a simulated "infinite bus," while the other was controlled by TPC. Training data was generated by applying white-noise inputs to the system. The evaluation involved step changes in active power reference.
- Grid-Connected Inverter Experiment: This experiment utilized a 25 kVA Egston COMPISO inverter connected to the actual Munich electric grid. Training data was collected from this setup. The evaluation involved step changes in active and reactive power references under real grid conditions.
No specific public datasets were introduced or used; rather, data was generated from the experimental setups themselves.
Key Insights and Ideas
- TPC is computationally feasible for real-time control: The most significant insight is that TPC, despite being a predictive control method, can run efficiently on standard microcontrollers at high frequencies (100 Hz in the two-converter experiment, and the control loop runs at 5 kHz with TPC operating at 100 Hz in the grid-connected experiment). This directly addresses the computational burden that plagues other data-driven predictive control methods.
- Model-free control works in practice for inverters: The experiments demonstrate that TPC can effectively control an inverter connected to a real, non-linear, and time-varying grid without needing an explicit grid model. This validates the "plug-and-play" potential.
- Output current constraints are critical and implementable: The paper successfully shows how to incorporate hardware current limits into the TPC framework using an SOCP formulation, ensuring safe operation.
- The grid can be "LTI-enough" for TPC: While the grid is inherently non-LTI, the experiments suggest that for the tested scenarios, the grid dynamics were sufficiently predictable for TPC to perform well, supporting the hypothesis that TPC can handle such systems.
Key Results
- Computational Performance:
- On a microcontroller, TPC achieved a solve time of 3 ms, enabling 100 Hz operation. In contrast, DeePC, even with reduced data, struggled to reach 10 Hz and exceeded memory limits.
- On a laptop, TPC consistently solved in 70 µs, while DeePC's solve time increased significantly with data size, reaching 40 ms with 500 data points.
- Two-Converter Experiment: TPC successfully tracked active power setpoints and enforced a 0.2 p.u. current magnitude constraint, demonstrating its ability to manage output limitations.
- Grid-Connected Experiment: TPC effectively tracked active and reactive power setpoints (0 to 0.8 p.u. for P, and 0.1 to 0.4 p.u. for Q) when connected to the Munich grid, with only minor tracking errors attributed to non-linearities at low reactive power setpoints.
Limitations and Future Work
- Grid Dynamics: The authors acknowledge that while the Munich grid was "LTI-enough" for their experiments, fully confirming TPC's effectiveness requires testing on inverters with more diverse grid dynamics, line impedances, and phase imbalances.
- Inverter Size: The grid-connected experiments were conducted on a 25 kVA inverter. Testing TPC on larger inverters is suggested.
- Scenario Complexity: Future work should include testing TPC in more complex grid scenarios, such as unbalanced three-phase conditions and fault scenarios.
- TPC Theory Extension: The current TPC theory is established for LTI systems. While it performed well on a non-LTI system, further theoretical exploration of its behavior in such environments would be beneficial.
Summary updated