Bozhen Jiang

Interpretability in short-term load forecasting (STLF) is crucial for power system operations, as it can earn the trust of power system dispatchers, align forecasts with operational practices, and support accountable decisions. Kolmogorov–Arnold Networks (KANs) have gained attention in this setting because of their strong interpretability and symbolic regression capabilities. However, complex black-box artificial neural networks (ANNs) may match or surpass KANs in predictive accuracy, which remains a central concern for dispatchers. To address this trade-off, we introduce a hybrid KAN–ANN model that improves accuracy while retaining interpretability. The model uses residual connections to enforce an additive decomposition of the load, and incorporates an L1 regularization penalty to ensure that the KAN component makes the dominant contribution, thereby maintaining interpretability. We further apply an alternating direction method of multipliers training scheme that switches optimization between the interpretable and residual parts, shrinking the non–interpretable component. In this setup, the KAN captures periodic and nonlinear structure, and the remaining variation is modeled by the ANN. We also visualize the alternating optimization to make the interaction between components clear. We evaluate the method against a state-of-the-art ANN–based STLF model using a public Swiss dataset and a large private dataset from Guiyang, China. Results show that the hybrid KAN–ANN approach improves KAN prediction performance while delivering accurate analytical expressions for STLF. Code is available at: https://github.com/BozhenJIANG/ KAN-ANN-for-STLF.

The AC optimal power flow (ACOPF) problem is fundamental for ensuring the economic and secure operation of power systems. In recent years, artificial neural network (ANN)-based methods have emerged as promising approaches for learning the mapping from load inputs to optimal generation solutions. However, these methods often function as "black boxes," offering limited transparency into their intermediate decision-making processes and making it difficult to interpret the underlying reasoning of the models. To address these challenges, we propose a direction field learning method (DFLM) for the ACOPF solution space, based on score matching and diffusion models. The direction field provides explicit information about the optimization direction, enabling intuitive and visual interpretation of the direction field learned by the ANN. This, in turn, enhances the transparency and credibility of the proposed method. Case studies on the IEEE 24-bus and 118-bus systems demonstrate the feasibility of the approach, showing that DFLM achieves performance comparable to state-of-the-art methods while offering significantly improved transparency.

The alternating current optimal power flow (ACOPF) problem can be solved efficiently and reliably using physics-informed neural network (PINN)-based solvers. However, load fluctuations and increases driven by economic development often cause significant distribution shifts, leading to degraded PINN performance. Therefore, timely updates of PINN parameters are essential to maintain accuracy in real-time applications. Conventional PINNbased supervised or unsupervised approaches, which rely heavily on labeled data or incur high costs due to gradient estimation, are often expensive and suffer from delayed responsiveness. To address these limitations, this paper proposes an unsupervised online learning framework for PINN-based ACOPF solver. First, a sequentially connected dual-network architecture is proposed, comprising an ACOPF solver network and an alternating current power flow (ACPF) evaluator network. This design establishes an explicit gradient propagation pathway between the two networks, ensuring full compatibility with existing deep learning frameworks. Second, physics-informed embedding schemes and corresponding loss functions are formulated for each network, along with derived gradient update rules. An unsupervised online learning method is further proposed, substantially reducing dependence on labeled data and enabling PINN online training. Additionally, an asynchronous training algorithm is introduced to balance the learning process by accommodating the differing convergence rates of the two networks. Numerical results on the IEEE 24-bus, 118-bus, and 1354-bus systems demonstrate that, compared to existing state-of-the-art PINN-based ACOPF solvers, the proposed method consistently produces nearoptimal solutions under data distribution shifts, highlighting its strong adaptability and stability.

Power flow (PF) calculation is essential for power system analysis. In recent years, data-driven methods have emerged as a promising approach to accelerate PF calculations. However, these methods require high-quality labeled data and often suffer from poor generalization. To address these issues, an unsupervised physics-informed neural network (UPINN) method is proposed for AC PF calculations. The proposed method follows the general process of Newton-Raphson’s method. By minimizing the physics-informed loss function, which is designed based on active and reactive power mismatches, the PF equations will be satisfied directly without the need to calculate the Jacobian matrix’s inverse. Proofs of the proposed UPINN training method’s convergence are provided. Case study results on the IEEE 24-bus and 118-bus systems demonstrate the feasibility of the proposed approach, showing that UPINN’s power flow model can achieve high generalization performance without relying on labeled data.

Forecasting load distribution provides valuable prior information for dispatching personnel, enhancing both the economy and security of power system operations. Despite the success of quantile and artificial neural networks (ANNs) based methods in short-term probabilistic load forecasting (SPLF), their blackbox nature limits their application and trust among power system operators. To address this issue, this paper proposes a novel interpretable SPLF approach based on Kolmogorov-Arnold Networks (KANs). This method incorporates the uncertainties associated with both input features and forecast loads. Forecast feature errors are modeled using a normal distribution to emulate input feature uncertainty, while forecast load uncertainty is estimated by approximating the residual sequence distribution derived from the training dataset. Additionally, we have developed a multi-interval linearized SPLF technique. By partitioning the range into intervals and using a truncated Taylor expansion within each sub-interval, the computation of load distribution is expedited. To validate the proposed approach, experiments are conducted by comparing the forecasting performance of KANs-based and quantile-based SPLF models using a publicly available dataset from Switzerland. Numerical results demonstrate the effectiveness of the proposed method in accurately forecasting load distribution. Our code can be found in https://github.com/BozhenJIANG/KAN-for-SPLF

The rapid rise in electric vehicle (EV) adoption demands precise charging station load forecasting, challenged by long-sequence temporal dependencies and limited data in new facilities. This study proposes MIK-TST, a novel two-stage transfer learning framework integrating Mixer, Informer, and Kolmogorov-Arnold Networks (KAN). The Mixer fuses multi-source features, Informer captures long-range dependencies via ProbSparse attention, and KAN enhances nonlinear modeling with learnable activation functions. Pre-trained on extensive data and fine-tuned on limited target data, MIK-TST achieves 4% and 8% reductions in MAE and MSE, respectively, outperforming baselines on a dataset of 26 charging stations in Boulder, USA. This scalable solution enhances smart grid efficiency and supports sustainable EV infrastructure expansion.

Short-term load forecasting (STLF) is indispensable for the efficient and cost-effective operation of power systems. The Kolmogorov-Arnold network (KAN), which is grounded in the Kolmogorov-Arnold representation theorem, exhibits superior interpretability and symbolic regression capabilities when juxtaposed with artificial neural networks (ANNs). It offers analytically expressible load formulas, making it a subject of recent research in short-term load forecasting. Nonetheless, as the depth of ANNs escalates, their predictive prowess frequently equals or even outstrips that of the KAN. To address this issue, this paper proposes a novel STLF method through the utilization and combination of Kolmogorov-Arnold Networks (KANs) and ANNs based on the idea of residual decomposition. KANs are utilized to uncover fundamental periodic and nonlinear patterns in the load, while the remaining residual component is fitted using ANNs. A hybrid model combining KANs and ANNs with residual connections is designed, explicitly satisfying load additivity. Additionally, regularization penalties are incorporated to ensure that the predictions made by KANs dominate, meeting the requirements of residual decomposition. Furthermore, a training method for the hybrid KAN-ANN model is proposed, which refines the analytical expressions obtained by KANs alone and provides more accurate forecast results. To validate the proposed method, we compared the forecasting performance of KANs, multi-layer perceptrons, and XGBoost using a public Switzerland dataset. Numerical results demonstrate that the proposed hybrid KAN-ANN method not only enhances the forecasting performance of the KAN-based method but also yields an analytical expression for STLF. Our code can be found in https://github.com/BozhenJIANG/KAN-ANN-for-STLF .