Deep learning and quantum-enhanced predictive optimization in power electronics

2.1. Classical optimization

Classical methods such as GA, PSO, and DE have been applied for converter parameter tuning, inverter efficiency improvement, and fault tolerance. Although effective, these methods often face issues of slow convergence and local minimum [12, 14-15].

2.2. Artificial intelligence and machine learning

Artificial Intelligence and Machine Learning (AI/ML)-based approaches have gained popularity for adaptive control. Reinforcement Learning (RL) enables dynamic adjustment of converter switching in real-time, while NN and DL models are widely used for predictive current control and fault detection [16].

2.3. Hybrid methods

Hybrid frameworks combine the robustness of fuzzy logic with optimization algorithms (Fuzzy and PSO) or Artificial Neural Networks (ANN) with evolutionary search (ANN and GA), providing both adaptability and fast convergence in multi-objective optimization [17].

2.4. Quantum machine learning

QML represents a new paradigm. Algorithms, such as Quantum Approximate Optimization Algorithm (QAOA) and Variational Quantum Eigen Solvers (VQE) show promise in solving nonlinear, high-dimensional optimization problems faster than classical solvers [18-27].

The performance features of classical optimization methods and artificial intelligence-based ones, hybrid optimization strategies, and the suggested Deep-Learning Quantum Machine Learning (DL-QML) framework are compared in Table 1. Classical methods are good and valid but have the weakness of being slow to converge and having low adaptability when applied to the nonlinear operating conditions. Intelligence AI methods enhance the ability to predict but remain largely dependent on the quality of training data. Hybrid optimization techniques are stronger but more complicated in terms of computation. By taking the proposed DL-QML framework, predictive modeling and quantum-assisted optimization are integrated, leading to increased efficiency, faster convergence rates, and higher-dimensional converter optimization problems.

4.1. Optimization of converter losses in MATLAB/Simulink

The Deep Learning module is a predictive estimation of optimal converter switching parameters based on historical data and current operational figures. The DL predictor compares the nonlinear relationships between converter operating variables and system efficiency compared to classical optimization methods that employ cyclic search methods. The input feature-vector is presented in form:

where, $V_{in}$ represents converter input voltage, ${ I}_{load}$ means load current, $f_s$ is the switching frequency, $R_{on}$ is the MOSFET on resistance, $T$ stands for the operating temperature.

The predicted converter efficiency is expressed as $\widehat{\eta }=f_{DL}\left(X\right)$.

A Multi-Layer Perception (MLP) architecture with two hidden layers is used as follows:

– Hidden layer 1: 10 neurons.

– Hidden layer 2: 10 neurons.

– Activation: ReLU.

– Training: supervised regression.

This supervised regression is carried out on synthesized converter data based on MATLAB/Simulink models. The DL module has three significant functions:

– Anticipates close optimum switching conditions.

– Minimizes the dimensionality of search optimization.

– Converts the QML optimizer more quickly.

The DL predictor is, thus, a preprocessing step followed by quantum optimization refinement.

Dataset Preparation and Training Strategy: The deep learning predictor that was trained was made with the simulated models of converter models in MATLAB that involved a high degree of variation in the operating conditions of input voltage, load current, switching frequency and resistance of the device. There were 500 artificial operating samples which were powered up to reflect realistic operating conditions of EV converters. The data was further broken down into the following parts:

– 70 % training data.

– 15 % validation data.

– 15 % testing data.

The trained neural network was able to attain constant levels in regression performance and low prediction error that allowed effective starting of the optimization search space.

4.2. Quantum machine learning-based optimization

The Quantum Machine Learning Optimization layer takes the predictions of switching parameters produced by the DL layer and optimizes them with Variational Quantum Optimization Algorithms (VQOA). Two algorithms assisted by quantum are considered.

1) Quantum Approximate Optimization Algorithm (QAOA).

2) Variational Quantum Eigen (VQE) Solver.

These algorithms approximate the constrained nonlinear optimization problem: $\underset{x}{\min }{P}_{loss}\left(x\right)$, subject to $\eta \left(x\right)\ge$ 90 %, where the decision vector is defined as ${x=\{ f}_s, R_{on}, D\}$ and $D$ is the converter duty cycle.

Optimization Workflow: The QML optimization cycle consists of Hamiltonian encoding of converter loss function, setting variational circuit parameters, hybrid quantum-classical loop of optimization, and expectation-value evaluation iterative parameter update

The QML layer is more effective in global optimization of nonlinear optimization problems of high dimensions as compared with classical optimizers, like GA and PSO.

4.3. Hybrid optimization

Hybrid adaptation layer is a layer that incorporates the concepts of Reinforcement Learning (RL) and that of Fuzzy Logic Control (FLC) to facilitate real-time correction of controllers in case of uncertain operating conditions. This layer makes up disruptions due to:

– Load variation.

– Switching noise.

– Temperature drift.

– Converter parameter mismatch.

1) Reinforcement Learning Controller: The RL agent keeps on updating switching actions because of converter performance feedback through a reward function that is defined as:

Policy updates follow gradient-based optimization:

where, $\alpha$ represents learning rate.

The RL controller enhances response in the short term and guarantees adaptative adjustment of switching.

2) Fuzzy Logic Controller: The fuzzy controller deals with uncertainty in the linguistic variables as follows:

– Input: error in efficiency due to frequency change of switching.

– Output: switching correction signal

Example rule: If efficiency reduces and there is load increment then there was a slight increase in switching frequency. The hybrid RL-FLC structure ensures stable converter operation under dynamic load environments.

The intended Deep Learning-Quantum Machine Learning (DL-QML) predictive optimization framework architecture towards efficiency boost of power electronic converters is shown in Fig. 1. The architecture incorporates three levels of processing, i.e. Deep Learning (DL) prediction, Quantum Machine Learning (QML) optimization, and Hybrid Adaptation control layers. The DL predictor estimates the near-optimal switching parameters using the parameters of operation, and the quantum optimization layer is minimized using variational search methods.

The hybrid reinforcement learning and Fuzzy Logic (FL) adaptation layer does real-time correction whenever there are changes in loads. This hierarchical architecture provides both offline optimization accuracy and online adaptive stability that is required in electric vehicle and renewable converter applications.

Fig. 1Proposed DL with QML-based predictive optimization framework for power converters

Fig. 2 shows the switching frequency dependence of converter efficiency to various strategies of optimization. As may be seen, the classical methods of optimization, including GA and PSO, offer moderate efficiency gains, owing to a low global search. The AI-based methods, viz. ANN and reinforcement learning, show better performance on efficiency, via the learning of nonlinear relationships of parameters. However, the proposed DL-QML framework achieves the highest efficiency across the switching frequency range due to predictive parameter initialization and quantum-assisted optimization refinement. This improvement is remarkably important for the increased switching frequencies at which switching losses control converter performance.

The hybrid learning and fuzzy control mechanisms are proposed to be used sequentially by the proposed DL and QML predictive optimization framework, which entails estimating optimal switching parameters using deep learning regression, loss minimization using quantum assistance and adaptive tuning in real time, respectively. Such a hierarchical design makes it possible to optimize offline and operate online. This enhances rational continuity between: Framework → Modeling → Simulation.

Fig. 2Hybrid control structure blending Classical, AI/ML, and QML modules

6.1. Optimization of converter losses in MATLAB/Simulink

1) Set $V_{in}$= 48 V, $I_{in}=$ 5 A, $I_{out}=$ 4 A, $t_{sw} =$ 50×10^-9 s, $P_{out}=$ 100 W.

2) Define $f_s\in$[10 kHz, 100 kHz], $R_{on}\in$ [20 mΩ, 100 mΩ].

3) Initialize best_loss = ∞.

4) For each: $\left(f_s, R_{on}\right)$

– Compute $P_{tot} = 0.5V_{in}{I}_{in}{t}_s{f}_s+I_{out}^2{R}_{on}$.

– Compute the efficiency, $\eta =\frac{P_{out}}{P_{out}+P_{tot}}\times 100 \%$.

– If $\eta \ge$ 0.9 and$P_{tot}<$ best_loss, update the best values.

5) Output best_fs, best_Ron, best_loss, $\eta$.

Optimal output values are shown below and in Table III for various efficiency thresholds. A 3-D loss surface plot, that is, the total loss vs. the sampling frequency, $f_s$ and the on resistance, $R_{on}$ are shown in Fig. 3.

Optimal Switching Frequency = 10 kHz.

Optimal $R_{on}=$ 0.02 W.

Minimum Total Loss = 0.38 W.

Efficiency = 99.62 %.

Table 3 demonstrates the efficiency increase which has been simulated utilizing various optimization techniques. It is possible to note that the classical optimization algorithms can offer a certain level of efficiency increase, whereas AI-based algorithms can offer a lot more performance due to their ability to predict nonlinear behavior. Hybrid optimization techniques are also more efficient in terms of combining two or more learning strategies. The offered DL-QML scheme possesses the greatest efficiency gain of all the compared schemes because it has a predictive initialization, as well as a quantum-assisted global optimization capacity. These results confirm the usefulness of the recommended framework in intelligent converter parameter tuning.

Fig. 33-D surface plot of losses of the converter (total loss vs fs and Ron)

Fig. 3 demonstrates the three-dimensional surface plot of losses of the converter with respect to the switching frequency and on resistance of the MOSFET. The plot clearly demonstrates that loss due to complete converter loss varies with switching frequency because of increased switching transition energy; conductive loss varies with resistance in the ON state. This nonlinear approach to the isolation of switching and conduction loss is verified by the surface curvature and demonstrates the necessity of intelligent optimization methods instead of conventional analytical optimization. The DL-QML optimization problem is an efficient way of searching for this nonlinear loss surface to find optimal operating points that meet the efficiency criteria.

6.2. Hybrid PSO and QML for converter efficiency optimization

1) Train MLP Regressor on: $x=$ [(24,10,20e3), (24,12,40e3), (36, 8, 30e3), (48,5,50e3)], $y=$ [0.85, 0.89, 0.92, 0.95].

2) Predict efficiency for test_case = (30, 9, 25e3).

3) Initialize best_eff = −∞. 4.

4) For $i=$ 1 to 1000:

– Sample $V_{in}$ ∼ U(20,50), $I_{load}$ ∼ U(5,15), $f_s$ ∼ U(20e3, 80e3).

– eff ← model.predict([$V_{in}$, $I_{load}$, $f_s$]).

– If eff > best_eff, update best_eff and best_params.

5. Output best_params and best_eff.

Optimal output values are shown below:

– Predicted Efficiency (DL Model): 146.4628

– Quantum-Inspired Optimal Parameters: $V_{in}$= 21.87 V, $I_{load}$ = 10.09 A, $f_s$ = 79.8 kHz.

– Optimized Efficiency = 472.3924.

6.3. Plot convergence of optimization methods

1) Generate iterations = 1: 20.

2) Compute GA, PSO, RL, QML results using ($c$ / iterations) + 0.05 × rand, where $c=$ {1.0, 0.8, 0.6, 0.3}.

3) Initialize plot (8, 6) and plot all curves:

– GA – circle markers.

– PSO – square markers.

– RL – diamond markers.

– QML – triangle markers, bold.

4) Label axes: “Iterations”, “Cost Function (J)”.

5) Add legend, grid, and title “Convergence of Optimization Methods”.

6) Apply tight layout and display plot$.$

Maximizing the techniques investigated of optimization reveals the convergence behavior. This has been proven to be especially faster in converging than both the GA, PSO, and RL methods because predictive negative feedback is implemented by the deep learning module, and parameter search is efficiently accelerated by the quantum-assisted optimization layer.

Fig. 4 makes a comparison of the convergence behavior of various optimization algorithms in converter parameter tuning. Since classical optimization methods like GA and PSO are based on stochastic searches, more iterations are needed to get solutions that are in a steady state. Reinforcement learning also exhibits better convergence performance although it still needs policy stabilization time through exploration. On the contrary, the suggested DL-QML optimization model is much faster to converge due to an informed starting point offered by the deep learning predictor and the acceleration of the global search suggested by the quantum-assisted optimization layer. The high speed of convergence renders the proposed framework to adapt well to real-time adaptive converters.

The suggested DL-QML optimization model will enable off-line parameter optimization and on-line adaptive control. This may be implemented into embedded processors or digital signal controllers to perform the DL prediction phase in environments with rapid switching parameter estimation, and the quantum optimization phase may be implemented offline, or using cloud-aided computation environments. The reinforcement learning and the fuzzy adaptation layer hybrid works in real time and constantly adjusts switching parameter on converter operating feedback.

Fig. 4Convergence of optimization plots (GA, PSO, RL, and QML)

6.4. Computational complexity analysis

It was found that the computational complexity of the suggested DL-QML optimization framework was low compared to classical optimization algorithms like GA and PSO. Classical heuristic algorithms usually need big iterative search populations, which leads to increased computation cost of high-dimensional optimization applications. The suggested framework minimizes the computational space by initializing the predictive parameters with the deep learning module, which narrows the search space significantly before the quantum optimization is selected to further. This means that fewer optimization steps are needed to achieve convergence. Moreover, the quantum-classical hybrid march helps to explore nonlinear solution space effectively when contrasted with a classical strategy of a search. These are the traits that render the proposed approach to real-time adaptive converter control applications.