Xu Cai

Atle Rygg, Marta Molinas, Zhang Chen et al.

Representations of AC power systems by frequency dependent impedance equivalents is an emerging technique in the dynamic analysis of power systems including power electronic converters. The technique has been applied for decades in DC-power systems, and it was recently adopted to map the impedances in AC systems. Most of the work on AC systems can be categorized in two approaches. One is the analysis of the system in the \textit{dq}-domain, whereas the other applies harmonic linearization in the phase domain through symmetric components. Impedance models based on analytical calculations, numerical simulation and experimental studies have been previously developed and verified in both domains independently. The authors of previous studies discuss the advantages and disadvantages of each domain separately, but neither a rigorous comparison nor an attempt to bridge them has been conducted. The present paper attempts to close this gap by deriving the mathematical formulation that shows the equivalence between the \textit{dq}-domain and the sequence domain impedances. A modified form of the sequence domain impedance matrix is proposed, and with this definition the stability estimates obtained with the Generalized Nyquist Criterion (GNC) become equivalent in both domains. The second contribution of the paper is the definition of a \textit{Mirror Frequency Decoupled} (MFD) system. The analysis of MFD systems is less complex than that of non-MFD systems because the positive and negative sequences are decoupled. This paper shows that if a system is incorrectly assumed to be MFD, this will lead to an erroneous or ambiguous estimation of the equivalent impedance.

Chen Zhang, Marta Molinas, Xu Cai et al.

Transients of a grid-synchronized voltage source converter (VSC) are closely related to over- currents and voltages occurred under large disturbances (e.g. a grid fault). Previous analysis in evaluating these transients usually neglect the nonlinear control effects of a VSC (e.g. phase-locked-loop, PLL). Therefore, potential stability issues related with nonlinear dynamics cannot be revealed properly. This work aims to move further in this respect. To better analyze and gain more insights into the nonlinear properties, dynamical analysis of a grid-tied VSC is conducted by parts. Specifically, the nonlinear behaviors of VSC power control loop (PCL) are firstly analyzed, in which the dynamics of PLL are assumed steady. Then, the nonlinear behaviors of PLL-dominant dynamics are further explored in detail, where the PCL is assumed steady. In this case, frequency instability and the mechanisms behind it are revealed. At last, effects of PQ controller regulation as well as controller bandwidth on the frequency stability are discussed. All the analysis and conclusions are verified by time domain simulations in PSCAD/EMTDC, where a switching model of VSC is adopted.

Atle Rygg, Marta Molinas, Chen Zhang et al.

This paper provides a comparative analysis of impedance models for power electronic converters and systems for the purpose of stability investigations. Such models can be divided into either decoupled models or matrix models. A decoupled impedance model is highly appealing since the Single-Input-Single-Output (SISO) structure makes the analysis and result interpretation very simple. On the other hand, matrix impedance models are more accurate, and in some cases necessary. Previous works have applied various approximations to obtain decoupled models, and both the dq- and sequence domains have been used. This paper introduces the terms decoupled and semi-decoupled impedance models in order to have a clear classification of the available approximations. The accuracy of 4 decoupled impedance models are discussed based on the concept of Mirror Frequency Coupling (MFC). By definition the decoupled models based on sequence domain impedances will be exact for systems without MFC. In the general case, they are expected to be more accurate than the decoupled dq-impedance models. The paper defines a norm $ε$ to measure the degree of coupling in the impedance matrices. This norm equals the error in the eigenvalue loci between the matrix and semi-decoupled models. This can also be viewed as the error in the semi-decoupled Nyquist plot. An example case study consisting of a grid-connected VSC with current controller and PLL is used to compare the different methods. It is found that decoupled and semi-decoupled models in the dq-domain are only applicable in grids with very low X/R-ratio. Furthermore, it is concluded that the decoupled model in the sequence domain gives close to equal results as the semi-decoupled model.

Chen Zhang, Marta Molinas, Atle Rygg et al.

Impedance is an intuitive and efficient way for dynamic representation of power electronics devices. One of the evident strengths, when compared to other small-signal methods, is the natural association with circuit theory. This makes them possible to be connected through basic circuit laws. However, careful attention should be paid when making this association since the impedances obtained through linearization are local variables, often referred to locally defined reference frames. To allow the operations of these impedances using basic circuit laws, a unified reference has to be defined. Though this issue was properly addressed on the state-space models, a thorough analysis and a clarification regarding the unified impedances and stability effects are still missing. This paper aims to bridge this gap by introducing the Impedance Operator (IO) and associated properties to the development of impedance networks. First, the IO for both the AC coupled and AC/DC coupled systems are presented and verified through impedance measurements in PSCAD. Then, three types of impedance network-based stability criterions are presented along with a clarification on the consistency of stability conclusions. Finally, the Nyquist-based analysis is explored, regarding the sensitivity to partition points, to open the discussion on the identification of systems weak points.

Jing Lyu, Qiang Chen, Xu Cai et al.

The small-signal impedance modeling of modular multilevel converter (MMC) is the key for analyzing resonance and stability of MMC-based ac power electronics systems. MMC is a converter system with a typical multi-frequency response due to its significant steady-state harmonic components in the arm currents, capacitor voltages, and control signals. Therefore, traditional small-signal modeling methods for 2-level voltage-source converters (VSCs) cannot be directly applied to the MMC. In this paper, the harmonic state-space (HSS) modeling approach is introduced to characterize the harmonic coupling behavior of the MMC. On this basis, the small-signal impedance models of the MMC are developed according to the harmonic linearization principle, which can include all the steady-state harmonic effects of the state variables, leading to the accurate impedance models. Furthermore, in order to reveal the impact of the internal dynamics and closed-loop control on the small-signal impedance of the MMC, three cases are considered in this paper, i.e., open-loop control, ac voltage closed-loop control, and circulating current closed-loop control. Finally, the analytical impedance models are verified by both simulation and experimental results.

Chen Zhang, Xu Cai, Atle Rygg et al.

Impedance models are widely used in assessing small signal stability of grid-tied voltage source converter (VSC) systems. Recent research has proven that impedance models of grid-tied VSC in both dq and sequence domains are generally Multi-Input Multi-Output (MIMO) systems, and the generalized Nyquist criterion has to be applied for stability analysis to these MIMO systems. However, finding Single-Input and Single-Output (SISO) equivalents for this system is always appealing because of the simplicity and the convenience for physical interpretation when assessing the stability, compared to MIMO systems. This paper presents two types of SISO impedance models of grid-tied VSC system, one is derived from the strong grid assumption, and the other is from the closed-loop equivalence. The accuracy of these models is assessed with respect to the measured impedances in PSCAD/EMTDC, and their effects on the stability assessment were analyzed as well. It is proven that the accurate SISO model gives identical result as the MIMO (matrix-based) impedance model with respect to stability analysis. However, the reduced SISO model may lead to wrong results if the bandwidth of phase locked loop is large.

Chen Zhang, Xu Cai, Atle Rygg et al.

Impedance-based stability analysis is appealing in the case of Single-Input Single-Output (SISO) systems. However, in the case of grid-tied voltage source converter (VSC) systems, dq impedances of source and load (VSC) subsystems are typically Multi-Input Multi-Output (MIMO) systems in which case the Generalized Nyquist Criterion (GNC) is required for analyzing the closed loop stability, which increases the complexity of the analysis. This paper explores further the coupling between positive and negative sequence impedances, in particular the dependency and bindings between them. It shows that the couplings in each subsystem can be compounded into two non-coupled sequence impedances if the source and load subsystems are viewed as an integrated system instead of as two separate subsystems. Therefore, two decoupled SISO systems are obtained which are defined as Augmented Sequence Impedance Networks (ASIN). The stability analysis of the closed loop system is performed directly on the ASIN with the principal of Argument. Both numerical and time domain verifications are presented.

Zheng Fang, Lichuan Xiang, Xu Cai et al.

ControlNet offers a powerful way to guide diffusion-based generative models, yet most implementations rely on ad-hoc heuristics to choose which network blocks to control-an approach that varies unpredictably with different tasks. To address this gap, we propose FlexControl, a novel framework that copies all diffusion blocks during training and employs a trainable gating mechanism to dynamically select which blocks to activate at each denoising step. With introducing a computation-aware loss, we can encourage control blocks only to activate when it benefit the generation quality. By eliminating manual block selection, FlexControl enhances adaptability across diverse tasks and streamlines the design pipeline, with computation-aware training loss in an end-to-end training manner. Through comprehensive experiments on both UNet (e.g., SD1.5) and DiT (e.g., SD3.0), we show that our method outperforms existing ControlNet variants in certain key aspects of interest. As evidenced by both quantitative and qualitative evaluations, FlexControl preserves or enhances image fidelity while also reducing computational overhead by selectively activating the most relevant blocks. These results underscore the potential of a flexible, data-driven approach for controlled diffusion and open new avenues for efficient generative model design. The code will soon be available at https://github.com/Anonymousuuser/FlexControl.

Minghao Xu, Lichuan Xiang, Xu Cai et al.

In this work, we question the necessity of adaptive gradient methods for training deep neural networks. SGD-SaI is a simple yet effective enhancement to stochastic gradient descent with momentum (SGDM). SGD-SaI performs learning rate Scaling at Initialization (SaI) to distinct parameter groups, guided by their respective gradient signal-to-noise ratios (g-SNR). By adjusting learning rates without relying on adaptive second-order momentum, SGD-SaI helps prevent training imbalances from the very first iteration and cuts the optimizer's memory usage by half compared to AdamW. Despite its simplicity and efficiency, SGD-SaI consistently matches or outperforms AdamW in training a variety of Transformer-based tasks, effectively overcoming a long-standing challenge of using SGD for training Transformers. SGD-SaI excels in ImageNet-1K classification with Vision Transformers(ViT) and GPT-2 pretraining for large language models (LLMs, transformer decoder-only), demonstrating robustness to hyperparameter variations and practicality for diverse applications. We further tested its robustness on tasks like LoRA fine-tuning for LLMs and diffusion models, where it consistently outperforms state-of-the-art optimizers. From a memory efficiency perspective, SGD-SaI achieves substantial memory savings for optimizer states, reducing memory usage by 5.93 GB for GPT-2 (1.5B parameters) and 25.15 GB for Llama2-7B compared to AdamW in full-precision training settings.

Yanchuan Chang, Xu Cai, Christian S. Jensen et al.

Trajectory similarity is fundamental to many spatio-temporal data mining applications. Recent studies propose deep learning models to approximate conventional trajectory similarity measures, exploiting their fast inference time once trained. Although efficient inference has been reported, challenges remain in similarity approximation accuracy due to difficulties in trajectory granularity modeling and in exploiting similarity signals in the training data. To fill this gap, we propose TSMini, a highly effective trajectory similarity model with a sub-view modeling mechanism capable of learning multi-granularity trajectory patterns and a k nearest neighbor-based loss that guides TSMini to learn not only absolute similarity values between trajectories but also their relative similarity ranks. Together, these two innovations enable highly accurate trajectory similarity approximation. Experiments show that TSMini can outperform the state-of-the-art models by 22% in accuracy on average when learning trajectory similarity measures.

Xu Cai, Jonathan Scarlett

The optimization of black-box functions with noisy observations is a fundamental problem with widespread applications, and has been widely studied under the assumption that the function lies in a reproducing kernel Hilbert space (RKHS). This problem has been studied extensively in the stationary setting, and near-optimal regret bounds are known via developments in both upper and lower bounds. In this paper, we consider non-stationary scenarios, which are crucial for certain applications but are currently less well-understood. Specifically, we provide the first algorithm-independent lower bounds, where the time variations are subject satisfying a total variation budget according to some function norm. Under $\ell_{\infty}$-norm variations, our bounds are found to be close to an existing upper bound (Hong et al., 2023). Under RKHS norm variations, the upper and lower bounds are still reasonably close but with more of a gap, raising the interesting open question of whether non-minor improvements in the upper bound are possible.

Xu Cai, Yang Wu, Qianli Chen et al.

We present an ultra-efficient post-training method for shortcutting large-scale pre-trained flow matching diffusion models into efficient few-step samplers, enabled by novel velocity field self-distillation. While shortcutting in flow matching, originally introduced by shortcut models, offers flexible trajectory-skipping capabilities, it requires a specialized step-size embedding incompatible with existing models unless retraining from scratch$\unicode{x2013}$a process nearly as costly as pretraining itself. Our key contribution is thus imparting a more aggressive shortcut mechanism to standard flow matching models (e.g., Flux), leveraging a unique distillation principle that obviates the need for step-size embedding. Working on the velocity field rather than sample space and learning rapidly from self-guided distillation in an online manner, our approach trains efficiently, e.g., producing a 3-step Flux less than one A100 day. Beyond distillation, our method can be incorporated into the pretraining stage itself, yielding models that inherently learn efficient, few-step flows without compromising quality. This capability also enables, to our knowledge, the first few-shot distillation method (e.g., 10 text-image pairs) for dozen-billion-parameter diffusion models, delivering state-of-the-art performance at almost free cost.

Xu Cai, Jonathan Scarlett

In this paper, we study the problem of estimating the normalizing constant $\int e^{-λf(x)}dx$ through queries to the black-box function $f$, where $f$ belongs to a reproducing kernel Hilbert space (RKHS), and $λ$ is a problem parameter. We show that to estimate the normalizing constant within a small relative error, the level of difficulty depends on the value of $λ$: When $λ$ approaches zero, the problem is similar to Bayesian quadrature (BQ), while when $λ$ approaches infinity, the problem is similar to Bayesian optimization (BO). More generally, the problem varies between BQ and BO. We find that this pattern holds true even when the function evaluations are noisy, bringing new aspects to this topic. Our findings are supported by both algorithm-independent lower bounds and algorithmic upper bounds, as well as simulation studies conducted on a variety of benchmark functions.

Xu Cai, Chi Thanh Lam, Jonathan Scarlett

In this paper, we study error bounds for {\em Bayesian quadrature} (BQ), with an emphasis on noisy settings, randomized algorithms, and average-case performance measures. We seek to approximate the integral of functions in a {\em Reproducing Kernel Hilbert Space} (RKHS), particularly focusing on the Matérn-$ν$ and squared exponential (SE) kernels, with samples from the function potentially being corrupted by Gaussian noise. We provide a two-step meta-algorithm that serves as a general tool for relating the average-case quadrature error with the $L^2$-function approximation error. When specialized to the Matérn kernel, we recover an existing near-optimal error rate while avoiding the existing method of repeatedly sampling points. When specialized to other settings, we obtain new average-case results for settings including the SE kernel with noise and the Matérn kernel with misspecification. Finally, we present algorithm-independent lower bounds that have greater generality and/or give distinct proofs compared to existing ones.

Xu Cai, Selwyn Gomes, Jonathan Scarlett

In this paper, we study the problem of Gaussian process (GP) bandits under relaxed optimization criteria stating that any function value above a certain threshold is "good enough". On the theoretical side, we study various {\em lenient regret} notions in which all near-optimal actions incur zero penalty, and provide upper bounds on the lenient regret for GP-UCB and an elimination algorithm, circumventing the usual $O(\sqrt{T})$ term (with time horizon $T$) resulting from zooming extremely close towards the function maximum. In addition, we complement these upper bounds with algorithm-independent lower bounds. On the practical side, we consider the problem of finding a single "good action" according to a known pre-specified threshold, and introduce several good-action identification algorithms that exploit knowledge of the threshold. We experimentally find that such algorithms can often find a good action faster than standard optimization-based approaches.

Xu Cai, Jonathan Scarlett

In this paper, we consider algorithm-independent lower bounds for the problem of black-box optimization of functions having a bounded norm is some Reproducing Kernel Hilbert Space (RKHS), which can be viewed as a non-Bayesian Gaussian process bandit problem. In the standard noisy setting, we provide a novel proof technique for deriving lower bounds on the regret, with benefits including simplicity, versatility, and an improved dependence on the error probability. In a robust setting in which every sampled point may be perturbed by a suitably-constrained adversary, we provide a novel lower bound for deterministic strategies, demonstrating an inevitable joint dependence of the cumulative regret on the corruption level and the time horizon, in contrast with existing lower bounds that only characterize the individual dependencies. Furthermore, in a distinct robust setting in which the final point is perturbed by an adversary, we strengthen an existing lower bound that only holds for target success probabilities very close to one, by allowing for arbitrary success probabilities above $\frac{2}{3}$.

Xu Cai, Yang Wu, Guanbin Li et al.

FRAME (Filters, Random fields, And Maximum Entropy) is an energy-based descriptive model that synthesizes visual realism by capturing mutual patterns from structural input signals. The maximum likelihood estimation (MLE) is applied by default, yet conventionally causes the unstable training energy that wrecks the generated structures, which remains unexplained. In this paper, we provide a new theoretical insight to analyze FRAME, from a perspective of particle physics ascribing the weird phenomenon to KL-vanishing issue. In order to stabilize the energy dissipation, we propose an alternative Wasserstein distance in discrete time based on the conclusion that the Jordan-Kinderlehrer-Otto (JKO) discrete flow approximates KL discrete flow when the time step size tends to 0. Besides, this metric can still maintain the model's statistical consistency. Quantitative and qualitative experiments have been respectively conducted on several widely used datasets. The empirical studies have evidenced the effectiveness and superiority of our method.

Jing Lyu, Marta Molinas, Xu Cai

This paper presents the harmonic state space (HSS) modeling of a three-phase modular multilevel converter (MMC). MMC is a converter system with a typical multi-frequency response due to its significant harmonics in the arm currents, capacitor voltages, and control signals. These internal harmonic dynamics can have a great influence on the operation characteristics of MMC. However, the conventional modeling methods commonly used in two-level voltage-source converters (VSCs), where only the fundamental-frequency dynamic is considered, will lead to an inaccurate model that cannot accurately reflect the real dynamic characteristics of MMC. Therefore, the HSS modeling method, in which harmonics of state variables, inputs, and outputs are posed separately in a state-space form, is introduced in this paper to model the MMC in order to capture all the harmonics and the frequency couplings. The steady-state and small-signal dynamic HSS models of a three-phase MMC are developed, respectively. The validity of the developed HSS model of a three-phase MMC has been verified by the results from both the nonlinear time domain simulation model in MATLAB/Simulink and the laboratory prototype with 12 submodules per arm.

Atle Rygg, Marta Molinas, Eneko Unamuno et al.

Impedance-based stability analysis in the dq-domain is a widely applied method for power electronic dominated systems. An inconvenient property with this method is that impedance models are normally referred to their own local reference frame, and need to be recalculated when referring to a global reference frame in a given network. This letter presents a simple method for translating impedance sub-models within a complex network, from their own reference frames to any given point in the network. What distinguishes this method is that by using a simple rotational matrix, it only needs impedance models in their own local reference frames, to be translated to a global reference in the network. By way of this method, standard circuit analysis rules for series and parallel connection are applicable, as proven in the letter. The method is defined and validated for impedances in the dq and modified sequence domains, and it is shown that the dependency on reference frame is marginal in the latter. An additional finding from the application of this method is that components or subsystems with a certain symmetry property called Mirror Frequency Decoupling are invariant to the choice of reference frame.