Source-linked AI summary
Flatness-based Neural Network Control of DC-DC Converters
Kamakshi Tatkare, Ufuk Topcu, Brian Johnson
TL;DR
Many learning-based converter controllers do not exploit converter structure, leaving the policy search broader than needed. The paper uses differential flatness to generate trajectories and train a static neural state-to-duty map; simulations show close duty agreement and regulation under large input and load changes.
Problem
Many learning-based converter-control methods do not exploit converter structure and therefore search over unnecessarily broad policy classes.
Source: §I · p. 1Method
The framework uses flatness-based trajectories to create labeled state-to-duty data and trains a multilayer perceptron to approximate the control law as a static map.
Source: §I · p. 1, C. Learning-based Control · pp. 3–4Results
Simulations on buck, boost, and buck-boost converters show learned duty closely matching model-generated duty and regulation under large input and load changes.
Source: B. Model Inference · pp. 6–7, §VII · p. 7Takeaways & Limitations
The numerical studies demonstrate feasibility on averaged converter models that exhibit pure flatness or flatness by prolongation.
Source: §VII · p. 7The demonstrated scope is averaged converter models exhibiting pure flatness or flatness by prolongation; formal stability guarantees and real-time implementation remain future work.
Source: §VII · p. 7
Abstract
from arXiv · showhide
In this work, we present a simple method to engineer a small neural network to control power converters. Learning-based methods have emerged as a promising approach for power converter control over wide operating point variations. However, many of these approaches do not exploit the analytical structure of converter dynamics and, therefore, search over unnecessarily broad policy classes. We address this gap with a systematic framework that combines model-based trajectory generation with policy learning. First, we use differential flatness of averaged converter models to generate feasible state and constrained duty cycle trajectories where duty acts as a control effort signal. Then for policy learning, we use these model-generated trajectories to construct labeled state-to-duty data and train a multilayer perceptron to approximate the control law with a static map. We apply this framework across buck, boost, and buck-boost converters through topology-dependent flat output selection and a common learning pipeline. Simulations show close agreement between learned and model-generated duty trajectories and stable regulation under large synchronous input and load disturbances.
I. INTRODUCTION
The paper frames converter control as duty-ratio regulation and argues that system structure can narrow learning-based controller searches. It proposes flatness-based state-to-duty learning and establishes the averaged converter models and steady-state assumptions used in the analysis.
I. INTRODUCTION
- The control objective is to maintain load voltage despite disturbances, using duty ratio—the fraction of each switching period that the switch is on—as the control effort.Source: §I · p. 1
- Many learning-based controllers search over broader policy classes than needed, while other approaches can require costly label generation, online adaptation, or complex training.The paper argues that converter structure can narrow the controller search.Source: §I · p. 1
- Flatness provides a direct state-to-duty relation for generating labels and training a multilayer perceptron to approximate the control law.The stated contributions also include feasible, constrained, smooth trajectories and numerical validation under large input and load disturbances.Source: §I · p. 1
A. System Modeling
- The averaged converter model uses a state vector and topology-dependent matrices, with equilibrium defined by zero state derivative and initial conditions assumed to be steady state.The initial state and duty are coupled by an algebraic relation.Source: A. System Modeling · p. 2
B. Flatness-based Trajectory Generation
Differential flatness parametrizes converter states and inputs using a flat output and finitely many derivatives, enabling trajectory-based control design. The section contrasts this structure with classical controllers whose robustness, response shaping, or implementation can be limited.
B. Flatness-based Trajectory Generation
- A flat output trajectory contains enough information to reconstruct system states and inputs through algebraic relations involving finitely many derivatives.A candidate output is suitable when the resulting constrained internal dynamics are well defined and stable.Source: B. Flatness-based Trajectory Generation · p. 2
III. CLASSICAL METHODS
- Small-signal linear feedback can perform well near its design operating point, but changes in load resistance and input voltage motivate large-signal methods.Source: §III · p. 2
- Sliding-mode control can exhibit chattering, while energy-based feedback focuses on stability rather than direct output-voltage response shaping.The described energy-based law also needs adaptive control for robustness to input-voltage changes when output voltage is measured.Source: §III · p. 2
- The paper presents a flatness-informed small neural-network approach against prior neural controllers that use multiple networks or pseudolinearization.Source: §III · p. 2
IV. PROPOSED DESIGN METHOD
The design uses topology-dependent flat outputs to implement a learned feedback map for converter control. For changing operating conditions, it smooths parameter transitions to avoid impulsive derivatives and duty commands.
IV. PROPOSED DESIGN METHOD
- The feedback map uses topology-specific flat outputs: output voltage for buck, inductor current for boost, and a separately defined output for buck-boost.The controller inputs shown are iL, vo, R, and Vin, and a PWM converts policy u into duty.Source: §IV · p. 2, Figure 2 · p. 3
A. Disturbance Modeling
- Sharp operating-parameter steps can create spikes in reconstructed derivatives and computed duty, so the design replaces the unit step with a smooth transition.Source: A. Disturbance Modeling · p. 3
- The smooth transition yields a bounded derivative, and its sharpness parameter k trades a steeper change against a longer transition interval.The paper notes that real-time changes in R and Vin are typically not ideal steps and may be measured or estimated.Source: A. Disturbance Modeling · p. 3
B. Response Shaping
The response-shaping procedure uses a second-order reference model for the flat output. A critically damped template is selected for smooth settling and minimal overshoot while preserving flatness-based parameterization.
B. Response Shaping
- A second-order critically damped reference transfer function shapes the flat-output trajectory of the averaged second-order converters.Source: B. Response Shaping · p. 3
- The selected template targets smooth settling and minimal overshoot, with desired natural frequency ω0; other transfer-function performance metrics can follow the same procedure.Source: B. Response Shaping · p. 3
- A unit-step reference is passed through the selected transfer function to obtain the flat-output response in the Laplace domain and then in time.Source: B. Response Shaping · p. 3
- Using the canonical second-order form retains the explicit parameterized structure of differential flatness while applying classical linear-system intuition to the desired dynamics.Source: B. Response Shaping · p. 3
C. Learning-based Control
The controller uses flatness to represent converter states and duty from a flat output and its finite derivatives, then learns a static neural map from sampled state-to-duty examples.
C. Learning-based Control
- Flatness represents converter states and duty through smooth maps of the flat output and a finite number of its derivatives.For the second-order averaged models, the state is x = [iL, vo]^T and duty lies in [0, 1].Source: C. Learning-based Control · p. 3
- Training examples come from sampling operating scenarios and smooth references, then discretizing the flatness parametrization into labeled regressor-duty pairs.The finite dataset contains trajectories, states, disturbances, and control inputs.Source: C. Learning-based Control · pp. 3–4
- An MLP maps the regressor to a duty-ratio estimate, with its weights and biases fitted using mean squared error.The network output is constrained to [0, 1].Source: C. Learning-based Control · p. 4
- Online control requires one network forward pass, with saturation available to enforce duty feasibility.Source: C. Learning-based Control · p. 4
V. APPLICATION TO DC-DC CONVERTERS
The converter application selects topology-specific flat outputs and derives state and duty trajectories from averaged models. For the buck case, abrupt parameter steps can make the derived duty impulsive or discontinuous, motivating smooth disturbance models.
V. APPLICATION TO DC-DC CONVERTERS
- The buck, boost, and buck-boost models use selected flat outputs based on internal-dynamics stability to derive flat parametrizations.The converter descriptions assume ideal circuits and ignore parasitics.Source: §V · p. 4
A. Buck Converter
- For the buck converter, output voltage y = vo is the flat output used to express inductor current and duty in terms of y and its derivatives.Source: A. Buck Converter · p. 4
- A step change in load resistance causes an instantaneous inductor-current jump and an impulse in its derivative, which the derived duty inherits.Source: A. Buck Converter · p. 4
- A sharp input-voltage step instead produces a finite jump in duty and numerical stiffness, so the method models Vin(t) and R(t) as smooth time-varying parameters before solving for duty.Source: A. Buck Converter · p. 4
B. Boost
The boost-converter parametrization uses inductor current as its flat output and extends the construction to time-varying disturbances. Because the parametrization is not pure, the method uses prolonged actuation.
B. Boost
- The boost converter uses y = iL as its flat output because this choice leads to stable internal dynamics.The remaining state and input trajectories are expressed in terms of y and its derivatives.Source: B. Boost · p. 4
- The boost parametrization is not pure, so the control action evolves over time as a new system state through prolonged actuation.Source: B. Boost · p. 4
- For disturbance transitions, the method introduces ϕ(t) as a unit-gain critically damped response and uses it with time-varying I⋆(t) to construct the flat variable.The resulting second derivative combines derivatives of I⋆ and ϕ.Source: B. Boost · pp. 4–5
- The load-transition sharpness k controls how smoothly R(t) and Vin(t) vary before prolonged parameterization yields the control-action and voltage-state equations.Source: B. Boost · p. 5
C. Buck-Boost
The buck-boost derivation uses inductor current as its flat output to parameterize voltage and duty. As in the boost case, the non-pure parametrization uses prolonged actuation and extends to time-varying input-voltage and load disturbances.
C. Buck-Boost
- The buck-boost converter uses y = iL as its flat output to parameterize output voltage and duty from y and its derivatives.Source: C. Buck-Boost · p. 5
- Because this parametrization is not pure, the method uses prolonged actuation; abrupt R and Vin changes follow the boost-case treatment.The stated transition derivation assumes nonzero input-source voltage and load resistance.Source: C. Buck-Boost · p. 5
- Smooth time-varying disturbances are incorporated by expressing the flat variable through ϕ(t), then deriving the required derivatives of I⋆ from R and Vin.The resulting prolonged parametrization updates the control-action and voltage-state equations.Source: C. Buck-Boost · p. 5
A. Data-generation and Training Setup
The dataset is built from flatness-parameterized trajectories sampled across operating points, initial conditions, and disturbance profiles. The setup splits the data for training, testing, and validation, then trains converter-specific MLPs.
A. Data-generation and Training Setup
- Latin hypercube sampling covers state operating points, initial conditions, and disturbance profiles for smooth trajectory generation and state-to-duty data construction.Each pair is organized as (z, u), with z containing feedback states and operating conditions.Source: A. Data-generation and Training Setup · p. 6
- The dataset is split into training, testing, and validation sets, and the models train with batch size 32.Table III reports converter-specific activation functions, layer widths, and trainable-parameter counts.Source: A. Data-generation and Training Setup · p. 6
- Training uses Adam with the learning rate and early-stopping epoch specified in Table III.Implementations use PyTorch on an Intel Xeon W9-3495X CPU workstation with an NVIDIA RTX 6000 Ada GPU.Source: A. Data-generation and Training Setup · p. 6
B. Model Inference
Across buck, boost, and buck-boost converters, the learned duty closely matches the flatness-generated duty and regulation is maintained through large input and load changes. The results are numerical studies on averaged models, with formal stability guarantees and real-time implementation left for future work.
B. Model Inference
- On unseen trajectories, learned and flatness-generated duty ratios nearly coincide for buck, boost, and buck-boost operation.The reported small mismatch occurs for buck-boost operation that includes both step-up and step-down behavior.Source: B. Model Inference · pp. 6–7
- Smaller smoothing parameter k spreads the boost transition and smooths duty action, while larger k steepens the transition and increases transient peaking in inductor current and output voltage.Source: B. Model Inference · p. 7
- The closed-loop responses show states settling to new steady values after short transients under simultaneous input-voltage and load changes.Duty jitter corresponds to the MLP inference rate.Source: B. Model Inference · p. 7
VII. CONCLUSIONS
- The learned duty closely matches the model-generated duty and maintains regulation under large input and load changes across the studied converters.Closed-loop responses show quick duty adjustment and state settling after simultaneous changes in input voltage and load resistance.Source: §VII · p. 7, B. Model Inference · p. 7
- The demonstrated scope is averaged converter models with pure flatness or flatness by prolongation; formal stability guarantees and real-time implementation remain future work.Source: §VII · p. 7