We introduce a physics-informed neural network for power flow (PINN4PF) that effectively captures the nonlinear dynamics of large-scale modern power systems. The proposed neural network (NN) architecture consists of two important advancements in the training pipeline: (A) a double-head feed-forward NN that aligns with power flow (PF), including an activation function that adjusts to the net active and reactive power injections patterns, and (B) a physics-based loss function that partially incorporates power system topology information through a novel hidden function. The effectiveness of the proposed architecture is illustrated through 4-bus, 15-bus, 290-bus, and 2224-bus test systems and is evaluated against two baselines: a linear regression model (LR) and a black-box NN (MLP). The comparison is based on (i) generalization ability, (ii) robustness, (iii) impact of training dataset size on generalization ability, (iv) accuracy in approximating derived PF quantities (specifically line current, line active power, and line reactive power), and (v) scalability. Results demonstrate that PINN4PF outperforms both baselines across all test systems by up to two orders of magnitude, not only in terms of direct criteria, e.g., generalization ability, but also in terms of derived physical quantities.

This paper investigates the impact of adaptive activation functions on deep learning-based power flow analysis. Specifically, it compares four adaptive activation functions with state-of-the-art activation functions, i.e., ReLU, LeakyReLU, Sigmoid, and Tanh, in terms of loss function error, convergence speed, and learning process stability, using a real-world distribution network dataset. Results indicate that the proposed adaptive activation functions improve learning capability over state-of-the-art activation functions. Notably, adaptive ReLU activation shows the most improved learning process, with convergence speed up to twice as fast as ReLU. Adaptive Sigmoid and Tanh activation functions exhibit superior performance in terms of loss function error, outperforming ReLU and LeakyReLU by up to two orders of magnitude. Furthermore, the proposed adaptive activation functions lead to smoother and more stable learning processes, especially during early training, improving convergence. The practical implications of this study include the potential application of these adaptive activation functions in distribution network modeling, forecasting, and control, leading to more accurate and reliable power system operation.

We investigate the performance of different annealers for power flow analysis using adiabatic computing. The annealers include D-Wave's simulated annealer Neal, D-Wave's quantum-classical hybrid annealer, D-Wave's Advantages system (QA), Fujitsu's classical simulated annealer, and Fujitsu's digital annealer V3 (DA). We implement Quadratic Unconstrained Binary Optimization (QUBO) and Ising model formulations, with the latter offering finer control over complex voltage adjustments. Different test systems are experimented with to systematically evaluate the annealers. The evaluation is based on the accuracy, the annealer's capability to handle the decision variables, and the computational time needed. QA and DA show superior performance over classical annealers for our application. DA effectively manages larger test systems, whereas QA encounters difficulties embedding the problem graph onto the hardware graph because of the limited qubit connectivity. This constraint confines QA to the 14-bus system with the QUBO formulation and the 4-bus system with the Ising model formulation. The best performance is associated with different annealers across different test systems, which suggests that adjusting the threshold can improve precision if the compiler and annealer are capable of handling the number of variables involved.

Metamodels are developed and used for aerodynamic optimization of a ducted opening integrated into a high-rise building to maximize the amplification factor within the duct. The duct consists of a nozzle, a throat, and a diffuser. 211 high-resolution 3D RANS CFD simulations are performed to generate training and testing datasets. The space-filling design and Genetic algorithm are used for data sampling and optimization, respectively. The performance of five commonly-used metamodels is systematically investigated: Response Surface Methodology (RSM), Kriging (KG), Neural Network (NN), Support Vector Regression (SVR), and Genetic Aggregation Response Surface (GARS). The investigation is based on (i) detailed in-sample and out-of-sample evaluations of the metamodels, (ii) annual available power in the wind (P_available), and (iii) annual energy production (AEP) for a 3-bladed horizontal-axis wind turbine (HAWT) installed in the mid-throat for the optimum designs obtained by the metamodels. The results show that converging-diverging ducted openings can magnify the experienced wind speed by the turbine and enhance the available wind power. In addition, the use of different metamodels can lead to a variation of up to 153% in the estimated P_available. For a small dataset, crude yet still acceptable accuracy can be achieved for Genetic Aggregation Response Surface and Kriging at a very low computational time.