This paper compares two large language model (LLM) agent architectures for modifying EnergyPlus Input Data Files (IDFs). In the tool-augmented approach, the LLM invokes high-level, domain-specific functions encapsulating heating, ventilation, and air conditioning (HVAC) engineering knowledge and IDF manipulation logic. In the autonomous approach, the LLM directly orchestrates primitive operations - object creation, deletion, and field-level modification. Both architectures are implemented as a multi-agent workflow using the Model Context Protocol (MCP), combining retrieval-augmented access to the National Energy Code of Canada for Buildings (NECB) standards, a simulation agent for IDF editing and EnergyPlus execution, and a coordinating supervisor agent. Seven NECB 2020 reference models spanning apartment, restaurant, school, warehouse, and office archetypes were evaluated on three modification tasks of increasing complexity. For a simple task (window U-value modification), both approaches achieved 100% success and produced identical energy results within each building. For a medium task (lighting-power reduction), both approaches succeeded in all seven buildings, although autonomous execution generally required more tool calls and time. For a complex task involving conversion from packaged terminal air conditioner (PTAC) or packaged direct-expansion (DX) systems to variable refrigerant flow (VRF), the tool-augmented approach succeeded in all seven buildings, producing simulation-ready, NECB-compliant models with zero severe errors. The autonomous approach failed in every case: the three apartment attempts stopped during planning or produced incomplete, invalid models, and the four packaged-DX attempts did not pass structural validation and therefore did not advance to annual simulation. Autonomous execution required substantially more tokens and agent processing effort, with no successful complex-task outcome. These findings indicate that autonomous LLM agents are viable for low-interdependency edits, but reliability degrades sharply when tasks require many coordinated changes. A hybrid strategy is recommended: autonomous execution for simple modifications and validated domain-specific tools for complex, compliance-critical transformations.
The integration of solar photovoltaic (PV) systems into the energy grid presents several challenges for controlling energy systems, primarily due to their intermittent nature. To manage this intermittency, accurate short-term forecasts are required. This study proposes a Bayesian-optimized sequence-aligned hybrid Convolutional Neural Network and Gated Recurrent Unit (CNN–GRU) framework to predict PV power 15 min ahead using multimodal sky images (64×64×48) and historical PV time-series data (16 × 1). Tree-structured Parzen Estimator (TPE) optimization with Hyperband pruning is employed to systematically search the hyperparameter space. On a standardized 1-year benchmark dataset, the proposed model achieves an overall Root Mean Square Error (RMSE) of 2.419kW, a Mean Absolute Error (MAE) of 1.153kW, and a Forecast Skill (FS) of 18.96%. These results demonstrate a 3.36% improvement in RMSE, a 7.83% improvement in MAE, and a 17.40% relative improvement in Forecast Skill compared to the state-of-the-art (SOTA) model, while achieving an 80.0% reduction in model parameters.
Ultra-short-term wind power forecasting is essential for secure grid operation and efficient energy management. However, accurate forecasting remains challenging because wind power series usually exhibit strong nonlinearity, stochasticity, and rapid local variations. To address this issue, this study proposes an integrated forecasting framework that combines intensity-directed causal preprocessing with prior-guided temporal modeling. Specifically, an Intensity-Directed Multi-Scale Causal Moving Average (ID-CMA) method is developed for strictly causal online preprocessing. By adaptively fusing causal moving averages at different temporal scales according to local variation intensity, the proposed method constructs a trend-residual representation that helps suppress noise while preserving rapidly changing characteristics. On this basis, a change-intensity-aware prior derived from the preprocessing stage is incorporated into the self-attention mechanism as an additive bias, allowing the model to place more emphasis on highly varying intervals during temporal dependency modeling. Experimental results on real wind power datasets confirm the effectiveness of the proposed framework for ultra-short-term wind power forecasting, while also indicating that the proposed preprocessing method remains effective across different forecasting models.
Accurate day-ahead solar irradiance forecasting is essential for photovoltaic integration, reserve scheduling, and energy storage operation. However, operational Numerical Weather Prediction (NWP) products are commonly available at hourly resolution, which limits their ability to represent sub-hourly irradiance ramps and cloud-induced fluctuations relevant to power system operation. To address this limitation, this study proposes Shape-Prior conditional Flow Matching (SPFlow), a probabilistic temporal-downscaling framework that converts hourly NWP forecasts into 15-min irradiance trajectories. SPFlow first employs a deterministic SPTransformer to construct a physically informed high-resolution shape prior from coarse NWP forecasts, satellite-derived meteorological variables, clear-sky irradiance, and site information. Conditioned on this informative prior, a conditional flow-matching model probabilistically refines the deterministic trajectory to represent unresolved sub-hourly variability and forecast uncertainty. Experiments on the SURFRAD dataset show that SPFlow consistently outperforms statistical, deterministic deep-learning, and probabilistic generative baselines. It reduces the continuous ranked probability score by up to 10.9% relative to the strongest probabilistic baseline, improves sub-hourly ramp representation, produces prediction intervals with near-nominal empirical coverage, and demonstrates promising zero-shot transferability to unseen locations. These results indicate that combining physically informed shape-prior construction with conditional flow-based probabilistic refinement provides an effective approach for converting coarse NWP forecasts into operationally useful sub-hourly probabilistic solar forecasts.
This study investigates the transferability of recommender systems for photovoltaic (PV) adoption prediction from entity-level data to spatially aggregated household data, addressing a major limitation imposed by privacy regulations such as the GDPR. By demonstrating how recommendation systems can operate on publicly available geospatial and socio-economic datasets, the work establishes a scalable framework for high-resolution green technology adoption analysis and energy system planning. The proposed methodology extends a hierarchical clustering–based recommendation framework through the integration of a Local Outlier Factor (LOF) noise-reduction layer to mitigate the increased heterogeneity and information loss induced by spatial aggregation. The framework is evaluated on a large-scale geospatial dataset comprising 87 features and approximately 417k building entries. Cohen’s d, Wasserstein distance, and CCDF-based analyses are used to assess the bias and structural changes introduced into the data by the noise reduction process, alongside variations in clustering distance metrics and LOF hyperparameters.The results demonstrate that recommender systems can be successfully transferred to spatially aggregated data, although spatial aggregation substantially increases noise and reduces baseline predictive performance. Incorporating the LOF-based robustness layer improves recall by 10% to 80% while maintaining limited structural distortion, thereby identifying a practical trade-off between predictive performance and bias introduction. The findings provide a robust foundation for recommender systems in green technology adoption research and highlight the broader applicability of spatially aggregated recommendation frameworks in data-restricted domains.
Fine-grained data center load trajectories are indispensable for scheduling, energy-efficiency diagnosis, and digital-twin calibration. In practice, their availability is constrained by metering hardware, storage and retention policies, historical monitoring configurations, and legacy monitoring systems that were deployed before the value of high-frequency load data for reconstruction and digital-twin applications was fully recognized. As a result, a target site may provide only coarse load records and a subset of auxiliary variables such as power usage effectiveness (PUE), temperature, and humidity. These observations differ in variable coverage, sampling interval, and missingness, making standard supervised reconstruction insufficient for this setting. This paper addresses 15min load reconstruction under heterogeneous weak observations. The proposed framework learns transferable temporal shape priors from external workload traces, encodes target-site evidence by its role, granularity, and reliability, and uses conditional diffusion to reconstruct a complete 15min trajectory. Load observations are also mapped back to their original granularity during training, which encourages generated curves to remain consistent with visible coarse evidence. On the 1h→15min task, the method achieves a mean absolute percentage error (MAPE) of 4.740%, an observation-consistency mean absolute error of 0.533%, a physical violation rate of 0.168%, and a correlation coefficient of 0.924. Cross-site, few-shot, external-source, and missingness experiments show that transferable shape priors, reliability-aware evidence fusion, and observation consistency jointly improve accuracy and operational plausibility.
Federated reinforcement learning is a promising approach for distributed artificial intelligence in cyber–physical energy systems, but many existing methods rely on black-box deep policies that reduce interpretability and may converge unstably under non-identical operating conditions. This work proposes a federated discrete reinforcement learning architecture for collective pitch control of wind turbines. The method combines federated learning, tabular reinforcement learning, and structured state–action representations so that turbines collaboratively learn transparent state–action value tables instead of opaque neural policies. The architecture is evaluated with multiple wind turbines exposed to independent wind realizations generated from the same Weibull distribution. Several federated aggregation strategies are compared in terms of convergence behaviour, training stability, scalability, synchronization frequency, parameter sensitivity, and communication requirements. The simulations show that the proposed proximal discrete aggregation method reached stable operation after approximately 30 training episodes, whereas Federated Averaging and the non-federated controller required approximately 70 episodes. FedSGD required the longest training time, converging after approximately 450 episodes, and presented the largest fluctuations, as reflected by the 95% confidence intervals. Increasing the number of participating turbines improved the convergence trend. These results demonstrate that discrete federated reinforcement learning can provide an effective balance between interpretability, convergence performance, and distributed learning efficiency in energy control applications.
Accurate modeling of yawed turbine wakes is essential for wind farm operation and control. Analytical wake models are computationally efficient but rely on simplified assumptions, whereas high-fidelity simulations capture complex flows at a high computational cost. Purely data-driven surrogates can be effective when sufficient data are available, but they disregard governing physical laws. This study investigates the training of a physics-informed generative surrogate framework for wake modeling in which a Deep Convolutional Conditional Generative Adversarial Network (DC-cGAN) is regularized by Reynolds-averaged Navier-Stokes (RANS) residuals. We quantify how embedding these residuals affects adversarial training behavior and improves steady-state, two-dimensional hub-height wake predictions across inflow conditions and yaw variations, resulting in physically consistent and data-efficient surrogates. To support physical fidelity, the underlying RANS simulations used for data generation were first benchmarked against field observations. Subsequently, the surrogate was trained and validated against these simulation targets, confirming its ability to capture key wake characteristics in both near- and far-wake regions. Within the tested parameter range, and over six independent training runs, the proposed framework achieves median test set R2 scores of 0.987 for streamwise velocity and 0.994 for spanwise velocity, median test set mean absolute relative error of 4.605%, and a minimum inference time of 0.0085 s per turbine while outperforming other machine learning baselines in wake flow field prediction. Furthermore, integrating physics‐based regularization improves training stability, reduces prediction error variability, and maintains predictive performance under the reduced-data and out-of-distribution settings examined in this study.
Methanol and its oxygenated blends are credible low-carbon fuels for direct-injection engines, yet classical correlations predict their spray behaviour poorly across practical pressures. Penetration sets the risk of wall impingement and cone angle governs air entrainment, so the two must be predicted together. High-speed Schlieren imaging in a constant-volume vessel recorded sprays of pure methanol and of blends with methyl tert-butyl ether, 2-ethylhexyl nitrate or acetone at 1, 5 and 10 % by volume, at injection pressures of 2 to 10 MPa and ambient pressures of 4 to 200 kPa, spanning flash-boiling and non-flashing regimes. Averaging five repeats per condition gave 4320 time-resolved observations across 144 operating conditions. A multi-output backpropagation neural network predicted both descriptors simultaneously and was evaluated on entire withheld spray conditions rather than individual frames. On 21 unseen conditions the backpropagation model reached coefficients of determination of 0.994 for penetration and 0.989 for cone angle, with root-mean-square errors of 2.00 mm and 1.82 degrees, at or below measurement repeatability. A calibrated Hiroyasu-Arai correlation remained useful under non-flashing conditions but lost most of its skill under flash boiling, where the backpropagation model did not. Gradient-boosted trees predicted cone angle marginally better, although only the backpropagation model describes both descriptors at once. Optimising its weights with a genetic algorithm gave no gain in accuracy. Penetration was governed by injection and ambient pressure, cone angle by additive composition. Withholding an entire ambient pressure level or additive concentration drove the coefficient of determination below zero, defining the model's practical limits.
Cloud cover is a key factor influencing photovoltaic energy production; however, monitoring remains limited by the sparse deployment of ground-based instruments and by intermittent satellite coverage. This study presents a machine learning framework, called Cloud Cover Estimation from Solar Panel data (CCESP). We use standard inverter logs from existing photovoltaic systems as distributed weather sensors and recover a meteorological variable that inverters do not measure directly. Unlike simulation-based explorations or specially instrumented prototypes, we provide the first large-scale, multi-site demonstration on real operational photovoltaic data without requiring additional hardware. We evaluate CCESP on three commercial photovoltaic sites in the Middle East with nine years of inverter records, paired with cloud cover data from multiple weather providers. The ensemble-of-inverters approach, implemented with XGBoost and contextual features, yields a mean absolute error of 10.96% (≈±11% cloud cover) and a Spearman correlation of 0.75 across sites. This error is lower than the smallest disagreement between any two of the independent weather providers (mean absolute error ≈14.76%) and approaches the accuracy of ground-based optical instruments reported in prior studies. In a sliding-window evaluation, the error remains stable across seven test years. At an independent site in the United States, the Spearman correlation is 0.74, compared with 0.73 at the Middle East sites under the same training target. Widely deployed photovoltaic infrastructure can serve as a cost-free supplementary sensing network that complements satellite and ground-based systems, particularly in regions with limited meteorological coverage, and operators can use the estimated cloud cover in solar power forecasting and grid integration.
Accurate reconstruction of unsteady flow fields in turbomachinery is essential for aerodynamic optimization and condition monitoring in energy and propulsion systems, but time-resolved measurements are frequently limited by hardware sampling capability and by basis mismatch under variable operating conditions. This study proposes a multi-condition joint sparse reconstruction framework that integrates cross-domain data assimilation, a weighted cross-correlation criterion, and mixed l2,1-norm optimization. A universal joint POD basis is developed from flow-field correlations among different incidence angles to improve cross-condition representation, while the row-sparse coefficient formulation exploits shared coherent structures across operating states. The method is validated using time-resolved PIV data from a compressor cascade at five incidence angles from 0° to 10°. Results show that the proposed POD-based joint reconstruction markedly outperforms conventional reconstruction methods. At a compression ratio of 0.6, the mean squared error is reduced by about 89.6% compared with the conventional POD-based approach. For the five-condition envelope, the weighted joint basis captures common structures across incidence states, although larger errors occur at high incidence due to stronger separation and condition-specific vortical dynamics. The method remains robust to different sampling-frequency combinations once the compression ratio exceeds the information threshold. The proposed framework provides an effective physics-informed data assimilation strategy for high-fidelity unsteady flow-field reconstruction under sparse sampling and variable operating conditions.
Accurate forecasting of energy consumption is important for the efficient operation of power systems, with direct implications for operational costs, energy management, and system maintenance. Due to the availability of extensive high-resolution consumption data from smart meters, data-driven methods have been used for short-term and long-term forecasting. However, their comparative performance on real-world smart meter data is still not well studied. In this paper, we present an empirical benchmark of nine modern deep learning models for time-series forecasting, including linear, MLP-based, convolutional, and Transformer architectures. We evaluate these models on two publicly available smart meter datasets. Our analysis focuses on three factors that strongly affect forecasting performance: the length of historical input, the prediction horizon, and the choice of model architecture. We show that extending the historical context improves accuracy, but only up to a saturation point, after which additional input provides limited benefit. In contrast, accuracy decreases as the prediction horizon increases. We also investigate the trade-off between prediction accuracy and computational complexity, and assess the statistical significance and practical magnitude of performance differences across models. Our results show that deep learning models consistently outperform classical baselines, while lightweight architectures achieve relatively similar performance at significantly lower computational cost. Additionally, architectural differences only become meaningful at longer forecasting horizons and on more heterogeneous datasets. Finally, a subgroup analysis across geodemographic and household categories shows that model choice has limited impact for most population segments, with attention-based architectures appearing to offer an advantage on the most under-represented or behaviorally irregular groups, though this advantage is not statistically robust given the small size of the affected subgroups and does not generalize consistently across datasets. This benchmark provides practical guidance for selecting forecasting models in smart meter energy applications.
Sodium-ion batteries have emerged as a promising alternative to lithium-ion batteries because of the abundance of sodium resources. However, their development remains constrained by the relatively low capacity and insufficient structural stability of cathode materials such as NaCrO2, while conventional trial-and-error approaches explore only a small fraction of the available chemical space. This work develops an integrated generative artificial intelligence and first-principles workflow for identifying chromium-containing sodium cathode candidates. A property-conditioned MatterGen model was fine-tuned using approximately 12000 density functional theory-relaxed sodium-containing materials spanning several Na-M-M’-O chemical systems. The fine-tuned model generated 56 chemically plausible structures, which were relaxed using a graph neural network interatomic potential, assessed for novelty, and screened according to the local coordination environments of sodium and transition-metal sites. The screening reduced 25 unique candidates to 9 structures for first-principles validation. All shortlisted compounds exhibited negative formation energies relative to their elemental reference states, ranging from -2.747 to -0.773 eV/atom, although this criterion alone does not establish ground-state stability. Calculated average desodiation voltages ranged from 0.651 to 3.864 V. Combined analysis of Bader charge change, local magnetic moments, and electronic structures identified chromium-dominant cationic redox, secondary-transition-metal-dominant cationic redox, mixed cationic–anionic compensation, and oxygen-dominated charge compensation. The resulting structures extend beyond conventional layered sodium chromium oxide and include layered, disordered-rocksalt, and nonlayered frameworks. These results demonstrate the potential of integrating generative artificial intelligence, machine learning, and quantum-mechanical simulations to expand the design space for sodium-ion battery cathodes and provide a transferable framework for other materials-discovery applications.
Gear transmission systems in wind turbines and other rotating energy machinery operate under variable rotational speeds, where vibration responses are governed by speed-dependent kinematic frequency variations. This speed-induced transformation challenges the construction of condition-consistent vibration representations and undermines reliable AI-enabled condition monitoring. To address this limitation, this study proposes a mechanics-consistent speed-invariant autoencoder (SiAE) for unsupervised fault detection using only healthy-condition data. Instead of treating speed variation as a generic statistical distribution shift, the proposed framework embeds gearbox kinematics into representation learning. First, the squared envelope order spectrum is used to map vibration signals into the order domain and suppress proportional frequency migration caused by rotational speed changes. Second, cross-condition interpolation is performed after order-domain alignment to construct a continuous and physically coherent healthy manifold across discrete speeds. Third, a speed-conditioned reconstruction mechanism with latent regularization is introduced to separate structural health information from deterministic speed-induced variations. Unlike conventional autoencoder-based anomaly detectors and existing speed-normalized approaches, which treat rotational speed as a generic auxiliary variable, SiAE jointly enforces kinematic consistency at the representation, augmentation, and model levels. The physical validity of the order-domain augmentation strategy is examined through theoretical analysis and gearbox dynamic simulation. Experiments under steady-state, time-varying, and completely unseen speed conditions, evaluated under fixed-load settings so that the influence of speed-induced kinematic effects can be isolated, show that SiAE produces stable and threshold-consistent health indicators, enabling robust fault detection across diverse operating regimes. These results demonstrate that integrating mechanics-based signal representation with AI learning improves the reliability of condition monitoring for variable-speed energy equipment.
Conventional approaches to accurate modelling of heat pumps and compressors are predicated on the availability of high-fidelity, data-intensive models for each component. Statistical and machine learning (ML)-based approaches are often-used, well-performing simplification pathways for purely physics-based modelling. TabPFN, a prior-data-fitted tabular foundation model leveraging the transformer architecture, consistently outperforms other ML approaches. This study investigates the performance of TabPFN in the domain of building energy systems (BES) by benchmarking its performance when predicting performance maps of heat pumps and compressors against other state-of-the-art approaches (XGBoost, Gradient Boost, Random Forest, Polynomial Regression). We develop an advanced training strategy incorporating group indices that buttress the ability of TabPFN to learn the basic, inherent characteristics from one dataset and transfer them when predicting the target variables of a nearly unseen dataset across variants and manufacturers. TabPFN outperforms all other approaches in nearly all instances. In the extreme case where only a single point from the target variant is included in the training dataset, average deviation in predictions remained under 10%. Group indices help in generalising effectively across manufacturers: when trained primarily on one manufacturer’s data, TabPFN yields high prediction accuracy (RMSE COP ≤ 0.2 and RMSE Q˙h ≤ 0.75) with the additional inclusion of only two points from the target variant from a different manufacturer. TabPFN can accurately predict equipment performance maps with minimal data, enabling rapid modelling of new component variants and reducing measurement requirements in BES.
An accurate estimation of the state of health (SOH) underpins safe and optimized use of the battery system. Although compelling, data-driven SOH estimation models typically require large amounts of high-quality labeled cycling data, while in practice such labels are often sparse in both quantity and coverage. Therefore, in this work, we propose a degradation-aligned self-supervised learning (SSL) framework based on a convolutional neural network-gated recurrent unit (CNN-GRU) model, which learns aging-consistent representations from unlabeled data through a cycle-order ranking objective as the pretext task for pretraining, thereby enabling robust SOH estimation after fine-tuning on sparsely labeled data. Test results showcase that the proposed ranking-based SSL approach proves to endow the pretrained model with degradation awareness from unlabeled data, and after fine-tuning the model can carry out accurate, robust SOH estimation, even when only an extremely limited amount of 1% of unevenly distributed labeled training data is available, where the MAE of 1.718% and RMSE of 2.329% can be achieved on the test cell. In addition, in-depth analyses are presented regarding the influences of label distribution and cross-cell robustness. We believe this work could shed new light on label-efficient SOH estimation of lithium-ion batteries, addressing a practical need in battery management.
The increasing demand for electric vehicle charging stations poses challenges for electric grid planning. This article proposes a framework for forecasting 24-hour peak-load profiles for new direct current fast-charging stations, defined here as planned or proposed stations with known static metadata, such as capacity, number of charging points, and location, but no historical load observations. The method combines K-means clustering of existing stations’ hourly load patterns, random forest classification of cluster membership for a new station, and quantile random forest prediction of peak-load magnitudes. Leave-one-out cross-validation showed that the conservative (α=0.99) peak-load profiles successfully covered the majority of observed hourly peak loads for 93.4% of charging stations while remaining below theoretical maximum capacity. The median prediction (α=0.50) achieved the lowest average prediction error among stations (mean absolute error = 108 kW), illustrating the trade-off between prediction accuracy and planning conservatism enabled by the quantile-based framework. Small charging stations are more difficult to predict, as they can more easily approach full capacity. The framework supports risk-aware grid planning by enabling planners to choose prediction quantiles that balance overload risk against excess capacity allocation, analogous to the classical newsvendor trade-off. The main contribution of this paper is a novel framework that allows grid planners to explicitly select planning risk through prediction quantiles while forecasting peak-load profiles for charging stations without historical observations. The prediction problem is simplified by dividing the problem into two more manageable subproblems: predicting the shape of the load curve and predicting its peak magnitude.
Smart energy management systems are shifting from passive monitoring toward proactive and increasingly autonomous operation, which makes timely anomaly detection essential to their security and resilience. This survey reviews anomaly detection for these systems through an agent-driven lens, and its central clarification is to separate a detector, which maps signals to anomaly scores, from an agent, which closes a perception–reasoning–action loop by invoking detectors, reasoning over operational context, and acting within a safety envelope or escalating to human approval. Following a transparent review protocol, we synthesize 164 studies across smart grids, distributed energy resources, battery storage, and electric-vehicle systems. We organize the field with an operational taxonomy that links each system layer to its data sources, anomaly types, detection tasks, response actions, and evaluation metrics, and we compare detector families along latency, robustness, interpretability, and physical consistency. Grading the evidence on a four-level validation scale reveals a consistent gap: detection accuracy is well documented, whereas latency, computational cost, and field performance are reported unevenly, and federated and agent-based methods remain supported mainly by simulation. We formalize an agentic orchestration framework with explicit agent roles, human-in-the-loop approval, and edge-to-cloud deployment, and we set out a staged roadmap toward standardized benchmarks, digital-twin and hardware-in-the-loop validation, and real-time deployment. The review aims to help researchers and practitioners choose, evaluate, and responsibly deploy anomaly detection in energy operations.
Our paper analyses the use of Python for Power System Analysis (PyPSA) in renewable and applied energy systems research. It is doing so by mapping the research literature that cites Python for Power System Analysis (PyPSA) rather than power-system analysis literature in general. Using 533 Scopus-indexed publications covering the period 2018-2026 that cite the original PyPSA article, this study combines VOSviewer term-network mapping with interpretable, metadata-based natural language processing (NLP), including TF-IDF, non-negative matrix factorization, latent semantic analysis, k-means clustering, and keyword co-occurrence analysisOur results show that PyPSA has evolved from a power-system analysis package into a methodological reference point for applied energy research on renewable integration, transmission planning, storage, sector coupling, hydrogen, market design, and Net Zero transition pathways. PyPSA-based modelling is becoming relevant for optimization in multi-energy systems, energy networks, smart grids, data analytics, and decision-making for energy problems.The paper’s contribution lies in applying explainable, unsupervised NLP to an energy-modelling corpus and producing document-level topic assignments, quantitative topic shares, temporal signals and an uncertainty-focused synthesis beyond standard visual bibliometrics. VOSviewer network analysis reveals five overlapping knowledge areas: energy-system modelling and optimization; renewable generation, storage and emissions; open-source software and power-flow modelling; hydrogen, decarbonization and sector coupling; and markets, dispatch and flexibility. The NLP workflow is consistent with these areas and identifies seven document-level topics. Our results show that PyPSA's contribution lies in connecting engineering detail with long-term optimization and in providing a common modelling logic for renewable generation, transmission, storage, hydrogen, heat, mobility and market design. Moreover, they suggest uncertainty and reporting checklists for future PyPSA-based studies and identify full-text AI-assisted evidence extraction as a priority. Our implications can be used in transparent modelling practice, future full-text NLP studies, and applied energy transition research.
High renewable penetration is reshaping the operation of distribution networks by increasing the variability, uncertainty, and spatial dispersion of power injections. In particular, photovoltaic (PV) generation introduces rapid voltage fluctuations, while electric vehicles (EVs) create flexible but highly time-varying demand. These changes fundamentally challenge conventional voltage regulation mechanisms in modern distribution systems. In modern smart grids, the integration of distributed energy resources (DERs), such as PV systems and EVs, has introduced new challenges in voltage regulation and power quality. As the penetration of inverter-based resources increases, maintaining voltage stability becomes more complex due to the variability and dynamic behavior of these resources. Traditional voltage–var control (VVC) methods, including centralized optimization, local droop control, and distributed schemes, are often insufficient to address the fast-evolving and decentralized nature of such systems. This paper proposes a two-stage coordinated VVC framework that integrates centralized day-ahead optimization with physics-informed AI-based local controllers for EV inverters in active distribution networks. Unlike conventional droop-based methods that rely on predefined low-dimensional control curves and centralized optimization approaches that require intensive communication and computation, the proposed framework enables fast local voltage regulation while preserving grid-level coordination. In Stage I, an AC branch-flow-based optimization coordinates OLTCs, capacitor banks, and PV/EV inverter set-points to obtain voltage-secure operating schedules. In Stage II, a residual multi-layer perceptron controller embedded in each EV inverter learns nonlinear reactive power correction policies from OPF-based solutions using only local measurements and scheduling information. Instead of conventional low-dimensional droop curves (Q(V), Q(P)), each EV inverter implements a residual multi-layer perceptron that maps rich local feature vectors—including voltage, active power, state-of-charge, tariff signals, and day-ahead set-points—to reactive power corrections. The network output is parameterized to satisfy inverter apparent-power limits by construction. The controller is trained offline using optimal set-points obtained from day-ahead or real-time optimal power flow, augmented with a physics-informed voltage penalty derived from the power-flow equations and admissible voltage bounds.
